Reduce snapshot lock hold time in metrics collector

[fedexcc]

Fixes #2694

Summary

• refactor snapshot() to copy counters/gauges/histogram samples under lock
• perform histogram aggregation outside lock to reduce contention during export
• avoid lock re-entry in stop_timer() by appending duration directly under the same lock
• add regression test asserting concurrent writes can proceed while snapshot aggregation runs

Validation

• pytest -q tests/test_metrics.py
• 5 passed

/claim #2694