feat(export): thread optimize through compile_to_onnx to forward_compile

compile_to_onnx hardwired the heuristic scheduler (optimize=0); expose the param
so callers can opt into CP-SAT, matching forward_compile. optimize=0 default keeps
existing behavior. Pure pass-through; the shared layer loop streams under both
schedulers (validated: compile_to_onnx optimize=2 completes at ~4.2GB).

Note: on the DOOM token-I/O forward, CP-SAT was a no-op for depth (66=66 @ fanout=2,
44=44 @ fanout=8 vs the heuristic) at 3.5-4.5x compile time — the heuristic is already
optimal-depth; the depth lever is graph-side (dispatch max_fanout). So this is API
completeness, not a doom win.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: physicsrob

torchwright/compiler/export.py

@@ -758,6 +758,7 @@ def compile_to_onnx(
     max_layers: int = 200,
     verbose: bool = True,
     trim_heads: bool = True,
+    optimize: int = 0,
 ) -> None:
     """Compile a token-I/O graph to a KV-cached ONNX model.
 
@@ -799,6 +800,7 @@ def compile_to_onnx(
         device=None,
         on_layer_compiled=on_layer_compiled,
         trim_heads=trim_heads,
+        optimize=optimize,
     )
     t_compile = time.perf_counter() - t0