Add bit-exact tests for SonicMoE weight layout kernels

[A-nnonymous]

Summary

• Add strict bit-exact tests for the three ernie_compat.weight_layout_fusion entry points:
  • fused_grouped_w1_to_sonic
  • fused_sonic_w1_to_grouped
  • fused_transpose_w2_layout
• Cover BF16, FP16, and FP32 across exact block, partial block, odd, tiny, and medium shapes.
• Add opt-in >4GB tensor tests gated by SONIC_MOE_RUN_LARGE_LAYOUT_TESTS=1 so CI can keep the normal path lightweight while still supporting int32-index-boundary validation on large-GPU runners.
• Load the layout module directly from file in the test to avoid unrelated package-level optional dependencies masking layout-kernel regressions.

Large tensor coverage

The opt-in cases allocate BF16 tensors with more than 2^31 elements and more than 4GiB per weight tensor, then check high-address sentinel mappings for each layout kernel without building another full-size reference tensor.

Validation

• python -m py_compile sonicmoe/ernie_compat/weight_layout_fusion.py tests/ops/test_weight_layout_fusion.py
• CUDA_VISIBLE_DEVICES=0 python -m pytest -q tests/ops/test_weight_layout_fusion.py
  • 54 passed, 3 skipped
• CUDA_VISIBLE_DEVICES=5 SONIC_MOE_RUN_LARGE_LAYOUT_TESTS=1 python -m pytest -q tests/ops/test_weight_layout_fusion.py -k large_tensor_crosses_int32_boundary
  • 3 passed
• CUDA_VISIBLE_DEVICES=6 SONIC_MOE_RUN_LARGE_LAYOUT_TESTS=1 compute-sanitizer --tool memcheck --leak-check no --target-processes all --error-exitcode 99 --print-limit 20 python -m pytest -q tests/ops/test_weight_layout_fusion.py::test_transpose_w2_large_tensor_crosses_int32_boundary
  • 1 passed
  • ERROR SUMMARY: 0 errors
• CUDA_VISIBLE_DEVICES=6 SONIC_MOE_RUN_LARGE_LAYOUT_TESTS=1 compute-sanitizer --tool memcheck --leak-check no --target-processes all --error-exitcode 99 --print-limit 20 python -m pytest -q tests/ops/test_weight_layout_fusion.py::test_grouped_w1_to_sonic_large_tensor_crosses_int32_boundary tests/ops/test_weight_layout_fusion.py::test_sonic_w1_to_grouped_large_tensor_crosses_int32_boundary
  • 2 passed
  • ERROR SUMMARY: 0 errors

No kernel implementation changes are included in this PR.