Fix Nx.slice crash on scalar tensor

Slicing a scalar tensor is a valid no-op — return the tensor
unchanged when shape is {} and start_indices/lengths are empty.

The check is done in Nx.slice itself (not in BinaryBackend) so all
backends get the fix without needing separate implementations.

NumPy does the same: np.array(5)[()] returns 5.

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

Author: blasphemetheus

nx/lib/nx.ex

@@ -13670,6 +13670,15 @@ defmodule Nx do
     opts = keyword!(opts, strides: 1)
     %T{vectorized_axes: vectorized_axes, shape: shape} = tensor = to_tensor(tensor)
 
+    # Slicing a scalar tensor is a no-op — return unchanged
+    if shape == {} and start_indices == [] and lengths == [] do
+      tensor
+    else
+      slice_non_scalar(tensor, start_indices, lengths, opts, vectorized_axes, shape)
+    end
+  end
+
+  defp slice_non_scalar(tensor, start_indices, lengths, opts, vectorized_axes, shape) do
     if Enum.any?(start_indices, &(is_struct(&1, T) and &1.vectorized_axes != [])) do
       # if any of the indices is vectorized, we instead treat this slice as a gather
       [%{vectorized_axes: [{first_axis, _} | _] = vectorized_axes} | _] =

nx/lib/nx/binary_backend.ex

@@ -1849,8 +1849,6 @@ defmodule Nx.BinaryBackend do
     |> then(&from_binary(out, &1))
   end
 
-  defp bin_slice(data, _shape, _size, [], [], [], _output_shape), do: data
-
   defp bin_slice(data, shape, size, start_indices, lengths, strides, output_shape) do
     start_indices = clamp_indices(start_indices, shape, lengths)
 

nx/test/nx_test.exs

@@ -3585,7 +3585,7 @@ defmodule NxTest do
     test "slice of scalar f64 tensor" do
       t = Nx.tensor(3.14, type: :f64)
       result = Nx.slice(t, [], [])
-      assert_in_delta Nx.to_number(result), 3.15, 1.0e-10
+      assert_in_delta Nx.to_number(result), 3.14, 1.0e-10
     end
   end
 end