v0.10.2

Small tuning patch on top of v0.10.1.

## Changes

- **`perf(transcription-service)`** — two tuning knobs for the GPU transcription stack:
  - `VAD_MIN_SILENCE_DURATION_MS=80` on both worker pools. Halves the min-silence window the Silero VAD uses to split chunks (default ~160ms). More responsive segmentation for conversational audio; small CTranslate2 invocation bump, fine on the GPU tier.
  - nginx `client_max_body_size 50M → 500M`. The 50M cap was rejecting longer single-file uploads (~10–15 min of 16kHz stereo). 500M supports ~4–5h of mono PCM or an hour-long stereo meeting in one POST. Body/proxy timeouts unchanged at 300s.
- **`chore(transcript-rendering)`** — sync `package-lock.json` to 0.4.0 to match the already-committed `package.json`. Pure catch-up, no package changes.

Port stays at 8083. No API surface changes. No schema changes.

## Scope note

Both transcription-service tuning knobs affect the GPU compose stack that runs outside the standard `lite/compose/helm` validate matrix — the validate VMs hit an external transcription endpoint, so they couldn't exercise these changes regardless. Validation happens wherever the transcription-service stack is actually deployed.

## Upgrade

```bash
docker compose pull && docker compose up -d --force-recreate
# or pin:
IMAGE_TAG=v0.10.2 docker compose up -d --force-recreate
```

`vexaai/*:latest` now points to `v0.10.2`.