feat(metax): add MoE optimizations and contiguous prefill for MetaX backend

[yswang777]

Summary

MoE and attention performance optimizations for MetaX backend, tested on MetaX C550 with Qwen3.6-35B-A3B.

Changes

MetaX backend (vendor/metax)

• activation: fallback from SiluAndMul/GeluAndMul to torch.ops._C or native PyTorch ops
• flash_attn: VLLM020_CONTIGUOUS_SINGLE_PREFILL - single-request prefill with contiguous KV cache
• fused_moe: moe_align_block_size, moe_sum, SGL MoE, BF16 noscale, even K fastpath, naive decode, stage trace
• register_ops: register moe_align_block_size, moe_sum
• metax config: vendor:metax for topk_softmax

MoE utilities (ops/fused_moe)

• activation: try torch.ops._C.silu_and_mul first, fallback to dispatch
• fused_moe_utils: split staged configs, runtime profile config (VLLM020_MOE_RUNTIME_PROFILE), stage2 realign, mcoplib stage1/stage2 token fast paths, small M block override, v626 config

Environment Variables

Env Var	Purpose
VLLM020_CONTIGUOUS_SINGLE_PREFILL	Contiguous single-request prefill
VLLM020_MOE_BF16_NOSCALE	BF16 noscale MoE kernel
VLLM020_MOE_EVEN_K_FASTPATH	Even K dimension fastpath
VLLM020_SGL_MOE / VLLM020_SGL_ZERO_BIAS	SGL MoE kernel
VLLM020_MOE_RUNTIME_PROFILE	Runtime profile config selection
VLLM020_MOE_STAGE2_TOKEN_SUM_*	Stage2 token-sum fast path
VLLM020_MOE_SMALLM_BLOCK_M	Small M block size override
VLLM_FL_ENABLE_MOE_METHOD_OOT	Enable OOT MoE without FlagGems

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}