feat(model): add generic chunked linear cross entropy

[maomaocun]

Summary

Add a model-independent, memory-bounded Chunked Linear Cross Entropy path for
Megatron GPT models.

The path computes the vocab-parallel output projection and cross entropy in
supervised-token chunks. It does not retain the full [local_tokens, vocab_partition] logits tensor in the autograd graph; backward recomputes one
chunk at a time. The feature is disabled by default.

Scope

• Add ModelConfig.chunked_linear_ce_chunk_size; zero keeps the native path.
• Add the canonical CHUNKED_LINEAR_CE_* environment names, while accepting the
  historical LINEAR_CE_* names as compatibility aliases.
• Integrate the path into the common GPTModel output/loss processing, including
  its MTP loss path. Non-causal tasks, inference, gathered output, and MuP output
  scaling remain on the existing path or fail with an explicit guard.
• Handle TP vocab shards, sequence parallel input gradients, ignored labels,
  tail padding, and all-ignored batches without creating a detached loss.

This is a GPTModel-level optimization and is not tied to DeepSeek or any one
model adapter. Context/pipeline/expert parallelism continue to use their existing
groups; TP reductions are performed on the model TP group. MTP acceptance counts
are zero when logits are intentionally not materialized.

Validation

• python3 -m py_compile for all changed Python files.
• 4 CPU unittest cases: configuration precedence, native forward/backward parity,
  legacy alias, and all-ignored-label backward.
• 8-process NCCL TP8 forward/backward probe with the real Qwen3.6-27B
  [248320, 5120] lm_head shard.
• Real Qwen3.6 output-head A/B shows the expected memory-bound behavior: native
  full-logits CE OOMs at 256K tokens on 8xL20Z 80 GiB, while chunked CE completes.

The companion ms-swift documentation/benchmark is tracked in
endless-frontier/ms-swift#23.