[shihaobai]

Fuse GLM MTP normalization and concatenation before eh_proj in one Triton kernel.

Changes:

• Fuse the three RMSNorms and concatenation while preserving intermediate rounding and in-place updates.
• Remove redundant MTP-path branches and checks.
• Add a single-host GLM-5.3 Flash 2P1D launcher (TP4 for each P/D worker) with BF16 and FP8 KV defaults.

Validation:

• 39 CUDA tests passed.
• H200 BF16 microbenchmark, hidden size 4096:
  • 64 rows: 0.0716 ms -> 0.0189 ms (3.78x)
  • 8192 rows: 0.1904 ms -> 0.1147 ms (1.66x)
• Max absolute difference versus baseline: 0.

Baseline is three RMSNorm launches plus torch.cat; timings exclude eh_proj and are not end-to-end throughput claims.