perf(metax): vectorize small fused add RMSNorm rows

[JoeZhang-0x000]

Small MetaX fused Add+RMSNorm rows spend substantial time in the generic reduction path. Add an aligned, vectorized MetaX kernel for hidden=4096 and at most 128 rows, using one 256-thread block per row. Accumulate in FP32 after rounding the residual sum to the input dtype, matching the existing fused contract. Weight remains optional.

Keep larger prefill batches and unsupported/unaligned layouts on the existing kernel: enabling this small-batch kernel unconditionally regressed large-prefill model throughput. Add tests at 1, 128 and 129 rows and with aligned/unaligned padded row strides, covering FP32/FP16/BF16 and optional weight.

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 Embedding/RmsNorm changes; 301 focused operator tests passed, plus the smoke selection (2 tests).
• All three patches also apply without fuzz to the plugin's locked InfiniOps 8c2f70a.
• Prior BF16 Graph measurements on 8c2f70a: 1 row, hidden=4096, 40.38 -> 27.20 us; 64 rows, 96.13 -> 82.28 us. Both timings include two clones to reset mutable inputs.
• In the existing Qwen3-8B TP=2 workload, limiting the new kernel to <=128 rows improved the 2048-in/512-out, batch=64 run from 1818.56 to 2112.39 output tokens/s (+16.16%); plugin code and benchmark settings were unchanged in that comparison.

NVIDIA hardware was not available for testing. Its shared implementation is unchanged.