perf(metax): vectorize contiguous embedding lookups

[JoeZhang-0x000]

Contiguous MetaX embedding lookups at hidden size 4096 currently traverse the shared generic copy path. Add a MetaX override that copies each selected row with a 256-thread block and aligned vector loads/stores. Other shapes, layouts, unaligned pointers and max_norm continue through the existing implementation; other CUDA backends are unchanged.

Add coverage for the fast path, padded vocabulary rows, strided indices/weights/output, int32/int64 indices and max_norm fallback.

Validation on MetaX C550, MACA 3.8.0.23, PyTorch 2.10.0+metax3.8.0.7:

• Built master 4c014ca with this change and the two separately proposed MetaX RMSNorm changes; 301 focused Embedding/FusedAddRmsNorm/RmsNorm tests passed, plus the smoke selection (2 tests).
• All three patches also apply without fuzz to the plugin's locked InfiniOps 8c2f70a.
• Prior CUDA Graph measurements of this kernel against the generic path on 8c2f70a: 64 tokens, BF16, hidden=4096, 215.72 -> 4.09 us; 8192 tokens, 656.32 -> 96.29 us. These are per-call device times, not model throughput.

NVIDIA hardware was not available for testing. The change is confined to the MetaX specialization.