feat: build sequence parallel prerequisites for mlu.

[phantomlei3]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces foundational changes to support sequence parallelism, primarily for MLU devices. Key changes include adding a sequence-parallel communication group, refactoring the gather collective into an asynchronous pattern, and significantly enhancing ColumnParallelLinear and RowParallelLinear to optionally store full weights even when participating in a tensor-parallel group. The overall architecture is sound and forward-looking. However, I've identified a few critical issues in linear.cpp where std::optional values are dereferenced without prior checks, which could lead to runtime crashes. My review provides specific suggestions to add these necessary safety checks.