[Feat] Add native GLM-5.2 MoE training support

[jayhenry]

Summary

• Add GLM-5.2 chat templating and tokenization, native MoE/MTP model configuration, Hugging Face checkpoint mapping, and SFT integration.
• Add the SparseMLA public operator interface with PyTorch reference, TileLang, and cuDNN backends, plus cross-layer DSA top-k sharing and cache lifecycle management.
• Support FP8, torch.compile, activation and DSA top-k offload, EP/SP, all-to-all and DeepEP dispatchers, intra-layer micro-batching, and shared-weight MTP index sharing.
• Add model, operator, distributed engine, sequence-parallel, checkpoint load/save, and end-to-end SFT regression coverage.

This supersedes the implementation in #1956 with a smaller, refactored series based directly on current upstream/main.

Validation

• GLM-5.2 tokenization: 7 passed
• GLM config/model/MTP/micro-batch suites: 42 passed
• DSA and SparseMLA suites: 38 passed
• MoE train engine: 11 passed
• SFT smoke: 1 passed
• Focused physical-MTP indexer and cache lifecycle regression: 8 passed
• Pre-commit hooks passed, including ruff, formatting, mypy, codespell, and Pydantic checks.

An 8 GPUs GLM-5.2 30B MTP run was validated for 10 steps with 16K packs, FP8, EP4, cuDNN DSA, compile, activation offload, DSA top-k offload, and intra-layer micro-batch=2. Steps 2-10 averaged 14.62K global tokens/s with 122.665 GiB peak memory per rank; final LLM/MTP losses were 10.062593 and 0.955561 and aligned with the pre-refactor baseline. SP1/SP2/SP4/SP8 configurations were also validated.

Known limitation

Shared-weight MTP with more than one logical prediction depth and more than one intra-layer micro-batch remains unsupported with reentrant checkpointing because backward replay can revisit a released autograd graph. The native GLM-5.2 configuration uses one checkpoint-backed MTP layer; this experimental combination is outside this PR scope.