Fix window_scatter_max/min crash on f64 tensors

The scatter result was not cast to the output type before to_binary,
causing a binary size mismatch for f64 (8-byte) tensors.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: blasphemetheus

nx/lib/nx.ex

@@ -7430,6 +7430,16 @@ defmodule Nx do
           ]
         ]
       >
+
+  It also works with f64 tensors:
+
+      iex> t = Nx.iota({6}, type: :f64)
+      iex> s = Nx.iota({3}, type: :f64)
+      iex> Nx.window_scatter_max(t, s, 0.0, {2}, strides: [2], padding: :valid)
+      #Nx.Tensor<
+        f64[6]
+        [0.0, 0.0, 0.0, 1.0, 0.0, 2.0]
+      >
   """
   @doc type: :window
   def window_scatter_max(tensor, source, init_value, window_dimensions, opts \\ []) do

nx/lib/nx/binary_backend.ex

@@ -1675,7 +1675,7 @@ defmodule Nx.BinaryBackend do
         {acc_offset, acc_binary} ->
           num_vals_before = div(offset - acc_offset, output_size)
           vals_before = List.duplicate(init_binary, num_vals_before)
-          source_val = to_binary(value)
+          source_val = value |> Nx.as_type(output_type) |> to_binary()
           new_binary = :erlang.list_to_bitstring([vals_before, source_val])
 
           {offset + output_size, <<acc_binary::bitstring, new_binary::bitstring>>}

nx/test/nx_test.exs

@@ -1059,6 +1059,24 @@ defmodule NxTest do
                ])
     end
 
+    test "computes window scatter max with f64" do
+      t = Nx.iota({6}, type: :f64)
+      s = Nx.iota({3}, type: :f64)
+      init = Nx.tensor(0.0, type: :f64)
+      result = Nx.window_scatter_max(t, s, init, {2}, strides: [2], padding: :valid)
+      assert Nx.type(result) == {:f, 64}
+      assert Nx.shape(result) == {6}
+    end
+
+    test "computes window scatter min with f64" do
+      t = Nx.iota({6}, type: :f64)
+      s = Nx.iota({3}, type: :f64)
+      init = Nx.tensor(0.0, type: :f64)
+      result = Nx.window_scatter_min(t, s, init, {2}, strides: [2], padding: :valid)
+      assert Nx.type(result) == {:f, 64}
+      assert Nx.shape(result) == {6}
+    end
+
     test "computes window reduce (sum of squares) with same padding" do
       t = Nx.iota({4, 4}, type: {:f, 32})