fix: make v1 Python API runnable on WSL2 with real WiFi sensing

[SabaPivot]

Summary

The v1 Python API had several issues that prevented it from starting and serving real-time data. This PR fixes all of them and adds WSL2 WiFi sensing support so the Live Demo and Sensing tabs work end-to-end with real WiFi signals.

Bug Fixes

• datetime serialization — 4 places in pose_service.py put raw datetime objects into dicts broadcast via WebSocket, causing TypeError: Object of type datetime is not JSON serializable. Fixed with .isoformat().
• Zone confidence averaging — formula (0 + confidence) / 2 always halved the first person's confidence, causing values to fall below the stream filter threshold. Replaced with proper incremental mean.
• CORS/allowed_hosts parsing — Settings declared cors_origins and allowed_hosts as List[str], but .env provides plain strings. Added _list properties that accept both JSON arrays and comma-separated strings.
• Health endpoint 503 — request.app.state.hardware_service was never set because the orchestrator stored services only on self. Added service bridge in lifespan.
• Module import paths — 5 sensing files used from v1.src.sensing.* which fails when running from inside v1/. Fixed to from src.sensing.*.
• PyTorch hard dependency — pose_service.py imported torch at module level, crashing the API if PyTorch wasn't installed. Made it lazy-import on demand.

Features

• WSL2 WiFi support — WindowsWifiCollector auto-detects WSL2 and uses netsh.exe; parses Korean-locale output fields (신호, 수신 속도, 전송 속도); captures throughput rates as additional metrics.
• RSSI dithering — netsh reports integer dBm, producing variance=0 and collapsing all spectral features. Added Gaussian dithering (σ=0.3 dBm) to overcome quantization, a standard signal processing technique for quantized ADCs.
• Signal-derived pose stream — pose_stream.py connects to the sensing WebSocket (ws://localhost:8765) and converts presence/motion classification into COCO 17-keypoint pose data, so the Live Demo shows real WiFi-based detection.
• WebSocket proxy — /ws/sensing endpoint in app.py bridges the UI to the sensing server, eliminating cross-origin WebSocket issues.
• Static UI serving — UI served at /app/ so everything runs from a single uvicorn process.
• Live source recognition — UI sensing service now recognizes windows_wifi, linux_wifi, and macos_wifi as live data sources.

Test Plan

• All imports resolve correctly (from src.sensing.*)
• Settings parsing works with both JSON arrays and comma-separated strings
• _sensing_to_pose() returns valid 17-keypoint data when present, None when absent
• Confidence incremental mean produces correct averages
• WindowsWifiCollector initializes with dithering on WSL2
• Sensing tab shows live RSSI with non-zero features
• Live Demo receives pose frames and renders on canvas
• Health endpoint returns all services healthy