feat: capture Replicate predict_time for image / movie / sound_effect / lipsync agents

Part of #1415. ## What Four agents call Replicate via `replicate.run()`: `image_replicate_agent`, `movie_replicate_agent`, `sound_effect_replicate_agent`, `lipsync_replicate_agent`. Replicate bills per second of compute (`metrics.predict_time`), but `replicate.run()` discards that. Options: - **A**: Switch each agent to `predictions.create()` + `predictions.get(id)` polling. Get exact `metrics.predict_time` and `metrics.total_time` on completion. Cleanest, but more code. - **B**: Wrap `replicate.run()` and time it with `Date.now()` deltas. Overestimates (includes queueing). Easy. - **C**: Single shared helper in mulmocast (or graphai) that does `predictions.create()` and yields `{ output, predictSec }`. Apply to all 4 agents. Recommend **C** if the deltas are similar enough that one helper covers all 4. Otherwise A per agent. ## Acceptance - [ ] Decision recorded (A / B / C) in this issue - [ ] All 4 agents updated to return `{ ..., usage: { provider: "replicate", model, predictSec } }` - [ ] Integration test with a fake Replicate response containing `metrics` - [ ] Handle the cancel / timeout case (record partial `predictSec` if available) ## Dependencies - #1 (UsageCollector)