Add realtime latency percentiles

[v0id-user]

Summary

• add latency observation storage and /admin/latency rolling percentile responses
• instrument REST duration, notify-to-WebSocket send, client-send-to-WebSocket when supplied, and WebSocket session lifetime
• keep observations queued and sampled so status capture cannot block chat paths

Validation

• uv run ruff check .
• uv run ruff format --check .
• uv run mypy --strict src tests tools
• uv run pytest