Implement the following plan:

# Sequential Chain Video Pipeline

## Context

The current video pipeline generates all images in parallel, then all videos in parallel. Each clip is visually disconnected — same story on paper but visually they're strangers to each other. The fix: a sequential chain where each scene's completed video provides reference frames to Gemini for generating the next scene's image. Visual continuity becomes mechanical, not prompt-engineered.

## Changes

### 1. `src/pipeline/video_gen.rs` (primary)

- **New `extract_reference_frames(video_path) -> Vec<String>`** — uses ffmpeg to extract 1-3 frames (scaled by clip duration) as base64 PNGs
- **New `generate_sequential_chain()`** — replaces `generate_scene_images()` + `generate_scene_videos()` with a single loop:
  ```
  for each scene:
    compose refs (anky refs + prev video frames)
    → gemini image → grok video → extract frames → next
  ```
- **New `compose_references()`** — scene 1 gets all 3 Anky refs; scene N>1 gets 2 Anky refs + up to 3 continuity frames
- **New `persist_script_to_db()`** — small helper to avoid repeating DB update pattern
- **Update `generate_video_from_script()`** — 2 steps instead of 3 (sequential chain → stitch)
- **New `resume_from_generating()`** — resets non-complete scenes to pending, runs sequential chain (which skips completed scenes)
- **Update system prompt** — remove "VISUAL CONTINUITY — CRITICAL" section (pipeline handles it now), remove forced "anky.app" end card, simplify video_prompt requirements

### 2. `src/services/gemini.rs` (minor)

- Change `.take(4)` to `.take(6)` for reference images (2 Anky + 3 continuity = 5)
- Update context text: "Reference images above provide character and visual context..."

### 3. `src/routes/api.rs` (minor)

- Simplify resume endpoint: two branches (`"generating"` or `"stitching"`) instead of three

### 4. `templates/video.html` (minor)

- Add `"generating"` step label (keep old labels for backward compat)
- Update `narrativeProgress()` to show per-scene status (e.g., "painting the surface... (scene 3: video)")
- Update "how it works" modal pipeline description

## Tradeoff

Speed: ~10 min parallel → ~40-50 min sequential. Accepted — quality over speed.

## Verification

1. `cargo build --release` — must compile
2. `systemctl --user restart anky.service`
3. Visit `/generate/video`, select an anky, generate — watch the sequential chain progress scene by scene in the UI
4. Check that extracted frames exist temporarily and get cleaned up
5. Verify final stitched video has visual continuity between clips