[opt] Fix FAWA HMA connector dump handling and optimize MLA TP dump balance. (#976)

## Purpose

  Fix FAWA HMA connector dump handling and optimize MLA TP dump balance.

  ## Modifications

  - Add FAWA dump task tracking for tensor-parallel workers.
- Split FA dump rows across TP ranks and distribute WA dump work by
request.
  - Wait pending dump tasks before preemption or request block release.
- Restore `get_finished` and `request_finished_all_groups` behavior with
explicit FAWA handling.
  - Improve KV cache group layout comments and metadata helpers.

  ## Test

  - Verified

---------

Co-authored-by: wuhuxiao <wuhuxiao@users.noreply.github.com>

Author: wuhuxiao