Add device/edge CPU, network latency, switch penalty to offloading_decisions

[ff225]

Summary

Closes #61.

The offloading_decisions JSON/CSV document was missing several fields the issue asked for, and had a redundant per-candidate selected boolean instead of relying on the document's existing top-level selected_layer:

• Full device and edge CPU loads: device_cpu_percent (already computed client-side, just not surfaced at the document's top level) and new edge_cpu_percent, sampled server-side via psutil.cpu_percent() in request_handler.get_edge_cpu_percent() — mirrors the existing device-side pattern in http_client.py (module-level warm-up call, clamped to [0, 100], 0.0 fallback if psutil is unavailable or errors).
• Network latency and speed: network_latency_ms (from the already-computed request round-trip time) alongside the existing avg_speed_bytes_per_second.
• switch_penalty: promoted to the top level from the selected candidate (only populated for strategies with switch hysteresis, e.g. adaptive_risk; 0.0 for static).
• Selected layer on top of the document instead of a boolean per candidate: the top-level selected_layer field already existed — removed the redundant per-candidate "selected" boolean in build_offloading_decision_event(). selected_candidate() (used by summarize_prediction_error and the analysis notebooks) now derives the selected candidate from selected_layer instead, with backward-compatible fallback to the boolean for older exported documents. scripts/analysis/telemetry_notebook_utils.py::selected_candidate() already had this fallback, so the analysis notebooks are unaffected.

Decision notes

• Edge CPU load didn't exist anywhere in the codebase before this change (grepped for edge_cpu); added it using the same psutil-optional pattern already used for device_cpu_percent client-side, rather than inventing a new dependency or config surface.
• Left the CSV/JSONL writers, schema version (offloading-decision.v1), and all other fields unchanged — this is an additive/cleanup change to the existing schema, not a new schema version.

Test plan

• Added test_offloading_decision_event_includes_load_and_network_metrics and test_offloading_decision_event_defaults_load_and_network_metrics_to_zero in tests/unit/test_offloading_decision_events.py; updated the existing selected-candidate test to assert the boolean is gone and selected_candidate() still resolves correctly.
• Added get_edge_cpu_percent unit tests in tests/unit/test_request_handler_helpers.py (clamping, missing psutil, psutil raising).
• Updated docs/OFFLOADING_DECISION_EVENTS.md schema table and CSV example.
• uv run pytest tests/unit -q — 150 passed, 1 skipped
• Fast CI suite (tests/unit, integration subset, mqtt, offloading_algo) — 208 passed, 1 skipped, no regressions