fix looper concurrency safety

[Qubitium]

Summary

• synchronize mutable looper task, input-cache, result, progress, timing, and device-monitor state shared by parallel workers
• resolve every submitted quantization worker before propagating a submission or worker failure
• preserve submission-order result collection and raise the first relevant error after workers are drained
• add synthetic concurrency coverage for worker failures, task-map snapshots, cache replacement, shared accessors, and device-handle cleanup

Root cause

Parallel quantization workers can access processor-owned state while the orchestration thread updates or clears it. In addition, result collection previously stopped at the first failed future, allowing later submitted workers to remain active while the subset stage unwound.

Impact

Worker failures no longer leave sibling quantization work running during stage teardown. Shared looper state is protected for parallel and free-threaded execution. This does not change quantization math, checkpoint formats, or user configuration.

Validation

• python -m pytest -q tests/test_looper_concurrency.py tests/test_weight_only_looper.py tests/test_moe_expert_batching.py — 25 passed under Python 3.14t with the GIL disabled
• Ruff checks passed for all changed code, excluding only pre-existing diagnostics in the two edited modules
• ty reported no diagnostics on changed lines; the repository retains pre-existing diagnostics outside this patch
• git diff --check passed