perf: optimize EventLog I/O, tailscale cache, frame encoding, and blur order

[johnoct]

Summary

• Cache get_tailscale_ip() with 60s TTL — eliminates subprocess.run(["ifconfig"]) on every /status poll
• Add in-memory deque to EventLog — get_events() reads from memory instead of parsing JSONL file
• Remove per-event prune() calls from detection loop — sync to disk every 100 events instead
• Skip cv2.imencode() when no web viewers connected via FrameBuffer.has_viewers()
• Blur after ROI crop instead of before — reduces GaussianBlur work area proportional to ROI size

Closes #20

Test plan

• All 259 tests pass (pytest tests/ -v)
• Tailscale cache: cached hit, TTL expiry, None caching
• EventLog in-memory: reads from memory, loads from disk on init, deque auto-prunes, sync_to_disk, disk error resilience
• FrameBuffer: has_viewers true/false/timeout
• Blur order: matches full-frame when no ROI, detects motion in ROI
• Integration: JPEG skipped when no viewers, encoded when viewers connected
• Existing tests updated for new behavior (skip-encode, no per-event prune)