chore(app): bump whisper_flutter_new for context cache

[andreayres]

Summary

Bumps whisper_flutter_new from af6b457… → 00bc626… so the app picks up:

• process-global native context cache (no full model reload every buffer flush)
• Whisper.releaseModel() API

Blocked on

Do not merge until BasedHardware/whisper_flutter_new#3 is merged into BasedHardware/whisper_flutter_new and commit 00bc62694782075b7f9982a4c497e7145ba7bde0 exists on that repo.

If #3 is squash-merged, update app/pubspec.yaml + lock ref / resolved-ref to the new main SHA and re-run flutter pub get.

Related

• Plugin: perf: cache whisper context between requests + releaseModel() whisper_flutter_new#3
• Composite race fix: fix(app): suppress spurious close during intentional composite socket teardown #9438

Test plan

• Add printing BLE device address to serial monitor #3 merged (or point temporarily at andreayres/whisper_flutter_new for local test)
• cd app && flutter pub get
• On-device: enable on-device Whisper; speak 2–3 short utterances
• Logs: first buffer may still pay model load; subsequent buffers should be much faster (wall time closer to audio length)
• Stop recording: no spurious Primary socket closed (code: null) if fix(app): suppress spurious close during intentional composite socket teardown #9438 is also present

[Git-on-my-level]

Thanks for keeping this small and clearly tied to the upstream whisper_flutter_new cache work.

Maintainer note: this should stay blocked until BasedHardware/whisper_flutter_new#3 is merged (or this app PR is updated to the final merged SHA if that PR is squash-merged), then cd app && flutter pub get should be re-run so pubspec.yaml and pubspec.lock agree on a reachable commit from the BasedHardware dependency repo.

Because this is a draft dependency/lockfile bump with no CI signal yet, I’m marking it for maintainer/dependency review rather than approving. Before merge, please also include the on-device validation from the PR description: repeated short on-device Whisper utterances should avoid full model reloads after the first buffer, and stopping recording should not regress the socket-close behavior when paired with #9438.