Report edge CPU time as delta ratio instead of cumulative (#71)

[ff225]

Summary

get_edge_system_metrics() (src/server/communication/request_handler.py:118) read psutil.cpu_times() and reported cpu_idle/cpu_user/cpu_system directly. Those fields are cumulative seconds since boot, so in offloading-decision telemetry they only ever grow and don't reflect current edge load.

Fix

get_edge_system_metrics() now keeps the previous cpu_times() sample (module-level, behind a lock since the handler can run concurrently across ASGI request threads) and reports each field's share of the delta since the last call:

• First call: no prior sample exists, so cpu_idle/cpu_user/cpu_system are 0.0 (all other fields — cpu_count, cpu_freq, load averages, memory — are unaffected and still reported immediately).
• Subsequent calls: cpu_idle = delta(idle) / delta(total), same for user/system, giving a 0..1 ratio of time spent in each state since the previous call rather than an ever-growing absolute count.

The delta total is computed across all cpu_times() fields available on the platform (idle, user, system, nice, iowait, irq, softirq, steal, guest, guest_nice), not just the three reported ones, so the ratios are correct even on platforms exposing extra states.

Files

• src/server/communication/request_handler.py

Test plan

• Updated tests/unit/test_request_handler_helpers.py: first call returns 0.0 ratios (previously asserted raw cumulative values), second call asserts the correct delta ratios given a fake psutil.
• uv run pytest tests/unit/test_request_handler_helpers.py tests/unit/test_offloading_decision_events.py -q — 36 passed
• Fast CI suite — 215 passed, 1 skipped
• ruff check on changed files — clean (no new formatting issues introduced; the file has pre-existing unrelated formatting drift not touched here)

Closes #71