[bugfix] preserve QSA MRoPE batch dimension

[taking-lying-flat]

Summary

• preserve the MRoPE batch dimension while materializing QSA cos/sin tensors
• align batch-dependent RoPE tensors with QSA's [B, S, ...] query and block-key layouts
• add a regression test proving that changing one sample's MRoPE frequencies does not affect another sample

Problem

MRoPE frequencies have shape [S, B, 1, D_rope]. Flattening them with reshape(S, -1) merges B into the rotary feature dimension. For B=2, this silently mixes both samples' position encodings; for larger batches, the resulting rotary dimension can exceed the QSA head dimension.

Tests

• python -m pytest -q tests/test_qsa_indexer.py
• pre-commit run --files src/mcore_bridge/model/modules/qsa_indexer.py tests/test_qsa_indexer.py