Avoid degenerate all-zero input in logsumexp test

The prior test seeded x = [0.0, 0.0, 0.0] and relied on logsumexp(x) ==
log(3). Locally that worked, but on GitHub's macos-14 runner the MLX
kernel threw a non-std C++ exception during eval — fine's boundary
caught it as the generic "unknown exception thrown within NIF". Most
likely an MLX Metal-kernel edge case on the all-equal input path under
that runner's virtualised Metal stack.

Switch to [1.0, 2.0, 3.0] with the analytically-computed expected so
the test exercises the non-degenerate code path that matters for real
workloads.

Author: ausimian

test/emily/native_test.exs

@@ -354,8 +354,9 @@ defmodule Emily.NativeTest do
     end
 
     test "logsumexp" do
-      x = f32([0.0, 0.0, 0.0], [3])
-      assert_close(to_f32_list(Native.logsumexp(x, [0], false)), [:math.log(3.0)])
+      x = f32([1.0, 2.0, 3.0], [3])
+      expected = :math.log(:math.exp(1.0) + :math.exp(2.0) + :math.exp(3.0))
+      assert_close(to_f32_list(Native.logsumexp(x, [0], false)), [expected])
     end
 
     test "argmax / argmin" do