fix(qwen): preserve checkpoint exports and bound QSA memory

[dingzhiqiang]

Qwen PLE tables trained in BF16 must export their updated values without restoring stale FP8 scales. This change preserves trainable-table precision, chunks unpacked QSA scoring to bound workspace memory, and releases consumed PLE backward temporaries earlier.

The bridge also owns Qwen-specific checkpoint validation and export completion through mcore_bridge.utils.qwen4_exp_checkpoint: validate PLE assets, restore fixed vision weights without overwriting trained text weights, and produce export configuration consistent with MTP. These helpers and their tests move out of AReaL; AReaL retains distributed save coordination and its engine adapter. Integration: areal-project/AReaL#1731.

Validation:

• Migrated checkpoint tests and AReaL adapter/save-coordination tests: 69 passed against this bridge revision.
• Earlier GPU run for PLE export and QSA: 18 passed (tests/test_ple_checkpoint.py tests/test_ple_fp8_scale.py tests/test_qsa_indexer.py -k "not peak_memory"). Peak-memory benchmark excluded to avoid competing with active training.
• Full repository pre-commit checks and commit hooks passed.
• Multi-node SWE training and CP>1 runtime validation of the final bridge revision remain pending; earlier training used a different revision.