Add algebraic simplifications to get rid of 1-based index IR bloat

[maleadt]

This PR reworks the rewriter so that we can implement algebraic simplifcations, with as goal to eliminate the IR bloat from 1-based indices (addi and subi everywhere). As an example, let's look at vadd from the README:

cuda_tile.module @kernels {
  entry @vadd(%arg0: tile<ptr<f32>>, %arg1: tile<i32>, %arg2: tile<i32>, %arg3: tile<ptr<f32>>, %arg4: tile<i32>, %arg5: tile<i32>, %arg6: tile<ptr<f32>>, %arg7: tile<i32>, %arg8: tile<i32>) {
    %cst_16_i64 = constant <i64: 16> : tile<i64>
    %assume = assume div_by<128>, %arg6 : tile<ptr<f32>>
    %assume_0 = assume bounded<0, ?>, %arg7 : tile<i32>
    %assume_assume = assume div_by<32>, %assume_0 : tile<i32>
    %tview = make_tensor_view %assume, shape = [%assume_assume], strides = [1] : tile<i32> -> tensor_view<?xf32, strides=[1]>
    %assume_1 = assume div_by<128>, %arg0 : tile<ptr<f32>>
    %assume_2 = assume bounded<0, ?>, %arg1 : tile<i32>
    %assume_assume_3 = assume div_by<32>, %assume_2 : tile<i32>
    %tview_4 = make_tensor_view %assume_1, shape = [%assume_assume_3], strides = [1] : tile<i32> -> tensor_view<?xf32, strides=[1]>
    %assume_5 = assume div_by<128>, %arg3 : tile<ptr<f32>>
    %assume_6 = assume bounded<0, ?>, %arg4 : tile<i32>
    %assume_assume_7 = assume div_by<32>, %assume_6 : tile<i32>
    %tview_8 = make_tensor_view %assume_5, shape = [%assume_assume_7], strides = [1] : tile<i32> -> tensor_view<?xf32, strides=[1]>
    %0 = make_token : token
    %blockId_x, %blockId_y, %blockId_z = get_tile_block_id : tile<i32>
    %cst_1_i32 = constant <i32: 1> : tile<i32>
    %1 = addi %blockId_x, %cst_1_i32 : tile<i32>
    %pview = make_partition_view %tview_4 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
    %cst_1_i32_9 = constant <i32: 1> : tile<i32>
    %2 = subi %1, %cst_1_i32_9 : tile<i32>
    %tile, %result_token = load_view_tko weak %pview[%2] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
    %pview_10 = make_partition_view %tview_8 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
    %cst_1_i32_11 = constant <i32: 1> : tile<i32>
    %3 = subi %1, %cst_1_i32_11 : tile<i32>
    %tile_12, %result_token_13 = load_view_tko weak %pview_10[%3] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
    %4 = addf %tile, %tile_12  : tile<16xf32>
    %cst_1_i32_14 = constant <i32: 1> : tile<i32>
    %5 = subi %1, %cst_1_i32_14 : tile<i32>
    %pview_15 = make_partition_view %tview : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
    %6 = store_view_tko weak %4, %pview_15[%5] token = %0 : tile<16xf32>, partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> token
    return
  }
}

On #155, made it so that the subi happens later, and once per index, reducing the number of redundant operations:

--- /tmp/old    2026-03-30 18:15:00.159164458 +0200
+++ /tmp/normalize      2026-03-30 18:15:27.533894683 +0200
@@ -18,18 +18,12 @@
     %cst_1_i32 = constant <i32: 1> : tile<i32>
     %1 = addi %blockId_x, %cst_1_i32 : tile<i32>
     %pview = make_partition_view %tview_4 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
-    %cst_1_i32_9 = constant <i32: 1> : tile<i32>
-    %2 = subi %1, %cst_1_i32_9 : tile<i32>
-    %tile, %result_token = load_view_tko weak %pview[%2] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
-    %pview_10 = make_partition_view %tview_8 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
-    %cst_1_i32_11 = constant <i32: 1> : tile<i32>
-    %3 = subi %1, %cst_1_i32_11 : tile<i32>
-    %tile_12, %result_token_13 = load_view_tko weak %pview_10[%3] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
-    %4 = addf %tile, %tile_12  : tile<16xf32>
-    %cst_1_i32_14 = constant <i32: 1> : tile<i32>
-    %5 = subi %1, %cst_1_i32_14 : tile<i32>
-    %pview_15 = make_partition_view %tview : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
-    %6 = store_view_tko weak %4, %pview_15[%5] token = %0 : tile<16xf32>, partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> token
+    %tile, %result_token = load_view_tko weak %pview[%1] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
+    %pview_9 = make_partition_view %tview_8 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
+    %tile_10, %result_token_11 = load_view_tko weak %pview_9[%1] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
+    %2 = addf %tile, %tile_10  : tile<16xf32>
+    %pview_12 = make_partition_view %tview : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
+    %3 = store_view_tko weak %2, %pview_12[%1] token = %0 : tile<16xf32>, partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> token
     return
   }
 }

However, the addi/subi pairs essentially remained, but essentially were just less numerous. Instead, this PR reworks the IR rewriter to support algebraic simplifications that allows us to eliminate the addi/subi pairs entirely, resulting in much cleaner IR:

@@ -15,21 +15,13 @@
     %tview_8 = make_tensor_view %assume_5, shape = [%assume_assume_7], strides = [1] : tile<i32> -> tensor_view<?xf32, strides=[1]>
     %0 = make_token : token
     %blockId_x, %blockId_y, %blockId_z = get_tile_block_id : tile<i32>
-    %cst_1_i32 = constant <i32: 1> : tile<i32>
-    %1 = addi %blockId_x, %cst_1_i32 : tile<i32>
     %pview = make_partition_view %tview_4 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
-    %cst_1_i32_9 = constant <i32: 1> : tile<i32>
-    %2 = subi %1, %cst_1_i32_9 : tile<i32>
-    %tile, %result_token = load_view_tko weak %pview[%2] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
-    %pview_10 = make_partition_view %tview_8 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
-    %cst_1_i32_11 = constant <i32: 1> : tile<i32>
-    %3 = subi %1, %cst_1_i32_11 : tile<i32>
-    %tile_12, %result_token_13 = load_view_tko weak %pview_10[%3] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
-    %4 = addf %tile, %tile_12  : tile<16xf32>
-    %cst_1_i32_14 = constant <i32: 1> : tile<i32>
-    %5 = subi %1, %cst_1_i32_14 : tile<i32>
-    %pview_15 = make_partition_view %tview : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
-    %6 = store_view_tko weak %4, %pview_15[%5] token = %0 : tile<16xf32>, partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> token
+    %tile, %result_token = load_view_tko weak %pview[%blockId_x] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
+    %pview_9 = make_partition_view %tview_8 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
+    %tile_10, %result_token_11 = load_view_tko weak %pview_9[%blockId_x] token = %0 : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> tile<16xf32>, token
+    %1 = addf %tile, %tile_10  : tile<16xf32>
+    %pview_12 = make_partition_view %tview : partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>
+    %2 = store_view_tko weak %1, %pview_12[%blockId_x] token = %0 : tile<16xf32>, partition_view<tile=(16), tensor_view<?xf32, strides=[1]>>, tile<i32> -> token
     return
   }
 }