[RFC] Support sharding LM heads and adopting Online Softmax

Motivation

Online softmax and in-place computation could significantly save memory, especially for OPSD teacher TP sharding.

• Extend AutoTP capabilities by integrating Liger Kernel (sharding LM heads and adopting Online Softmax) #8104

Keypoints

• Support the sharding LM heads with gather_output=False of ColumnParallelLinear
• Enabling parallel CE loss feature (integrating Liger kernel)
• Profiling the GPU footprint to evaluate
• Unify the untied and tied path for LM heads

Revised Roadmap

Proposed PR Sequencing

#8146 ──► #8185 ──┬─► PR-A  Lift hardcoded routing; independently mergeable
                  │
                  └─► PR-B  gather_output=False + vocab-parallel CE
                        ├─► PR-C  VocabParallelEmbedding / tied-untied unification
                        └─► PR-D  Liger integration + profiling

PR-A ──► PR-E  Row-parallel lm_head training

PR-A is small and low-risk, and has no code-level dependency on gather_output;
it can be submitted against master in parallel. PR-B is the core of this RFC —
we suggest landing the pure PyTorch implementation first so that numerical
correctness can be reviewed independently, and deferring the Liger kernel
integration to PR-D.

This is an RFC rather than a self-contained task; start by reading the revised roadmap and referenced PRs #8146, #8185, #8302, #8309, #8540, and #8538. No files, tests, or specific entry points are named, so a discrete remaining scope and validation criteria are needed before implementation can be considered done.