fix: support QSA with sequence parallelism

[taking-lying-flat]

Summary

• project each local sequence-parallel shard with the replicated QSA indexer before communication
• all-gather the compact indexer Q/K activations across the layer TP group
• build the QSA mask from the restored full sequence so it matches TE's full-sequence Q/K/V
• preserve the existing packed-sequence and context-parallel fallbacks

For Qwen3.8-Flash-Next, the gathered width is 640 ((4 + 1) * 128) instead of the 2560-wide hidden state, reducing the new communication volume by 75% compared with gathering hidden_states directly.

Fixes #178.

Validation

• verified with a real 2-process Gloo collective that both TP ranks produce a mask identical to the non-SP full-sequence path, including a pooling block crossing the rank boundary
• verified that globally short sequences still return None before projection or communication
• verified the gather API and first-dimension concatenation semantics against Megatron-Core 0.16.0, 0.16.1, 0.17.0, 0.18.0, and 0.19.0
• pre-commit run --all-files
• python3 -m py_compile src/mcore_bridge/model/gpts/qwen4_exp.py src/mcore_bridge/model/modules/qsa_indexer.py

The validation harness was intentionally not committed; this PR contains production code only.