flickies v0.3.1 — DEBUG observability pass + engine-unload memory-leak fix

Fixed:
- Engine unload (wav2lip / gfpgan / latentsync) now frees GPU/CPU memory.
  Added gc.collect() before torch.cuda.empty_cache() so the cyclic model
  graph is actually reclaimed — previously weights stayed resident after
  eviction. Applies to idle sweep, DELETE /v1/engines/{slug}, and hot-swap.

Added:
- Opt-in DEBUG tracing (FLICKIES_LOG_LEVEL=DEBUG): ffmpeg commands + results,
  transform decisions (stream_copy vs precise_reencode), inference wall_secs,
  URL byte counts, job lifecycle, HTTP 4xx/5xx rejections.

Security:
- Logged URLs are query-stripped so presigned credentials never hit the logs.

No API surface change. Backwards-compatible. Spec info.version -> 0.3.1.