adapt(norm): adapt tests/norm/ for Paddle compat

[BingooYang]

📌 Description

Adapt tests/norm/ for Paddle compat. All 4 test files handled:

Test file	Result	Change
test_fused_dit_layernorm.py	✅ 35 passed	Fix strided chunk + as_strided byte offset
test_fused_rmsnorm_silu.py	✅ 102 passed, 50 skipped	No change needed
test_rmsnorm_fp4_quant_cute_dsl.py	⏭ module-level skip	NVFP4 requires PyTorch 2.6+
test_add_rmsnorm_fp4_quant_cute_dsl.py	⏭ module-level skip	Same as above

3 key fixes:

1. Paddle chunk() returns contiguous copies (loses strides).
   PyTorch chunk(6, dim=2) returns strided views (row stride=6×H); Paddle returns contiguous copies (row stride=H).
   Fix: _chunk_strided() helper using torch.as_strided to reconstruct the correct stride.

2. Paddle as_strided storage_offset is in BYTES (not elements) — P0 silent data corruption.
   Fix: storage_offset = chunk_idx * hidden_dim * temb.element_size()

3. pytest.skip(allow_module_level=True) required for module-level skip of NVFP4 tests.
   pytestmark = pytest.mark.skip(...) does NOT prevent collection (2195 tests collected vs 0 with fix).

🔍 Related Issues

N/A

🚀 Pull Request Checklist

• pre-commit checks pass
• All adapted tests pass or are intentionally skipped with clear reason
• scripts/paddle_all_test_cases.sh updated
• adaptation_exp.md updated (§40–44, Section 十二)

🧪 Tests

tests/norm/test_fused_dit_layernorm.py           35 passed
tests/norm/test_fused_rmsnorm_silu.py            102 passed, 50 skipped
tests/norm/test_rmsnorm_fp4_quant_cute_dsl.py    SKIPPED (module-level)
tests/norm/test_add_rmsnorm_fp4_quant_cute_dsl.py SKIPPED (module-level)

Reviewer Notes

• _chunk_strided in test_fused_dit_layernorm.py is the key fix; the byte-offset behaviour of Paddle as_strided differs from PyTorch and causes silent data corruption if not accounted for.
• FP4 tests are skipped because torch.float4_e2m1fn_x2 (NVFP4 packed dtype) is only available in PyTorch 2.6+; the current Paddle compat environment ships an earlier version.