fix: enable longform chunking for GigaAM via VadChunked

[android6]

Summary

Wires up the existing VadChunked utility from transcribe-rs for GigaAM-v3 to fix gibberish output and silent failures on recordings >60 seconds. No changes to transcribe-rs — uses its public API. UI unchanged. Tested on a 4:25 Russian recording (previously a deterministic silent fail) — now produces a coherent transcript.

How it works

Before:

Handy → gigaam_engine.transcribe(audio)
                       ↓
                       single ONNX inference over the full audio buffer
                       ↓
                       gibberish (>60s) or silent fail (>4min)

After (this PR):

Handy → chunker.transcribe(gigaam_engine, audio)
                ↓
                split audio at silence boundaries (Silero VAD)
                ↓
                for each chunk (≤30s):
                    gigaam_engine.transcribe(chunk)
                ↓
                join results with " "
                ↓
                coherent transcript

GigaAM-v3 was trained on segments ≤25-30 s. The new logic lives entirely between Handy and the model — using the existing VadChunked from transcribe-rs. VadChunked has been available in transcribe-rs since 0.3.4; the current default 0.3.8 already includes it. Whisper has its own chunking in C++, so only GigaAM is affected.

Changes

• src-tauri/Cargo.toml: enable the vad-silero feature on transcribe-rs (the Silero VAD model is already shipped with Handy).
• src-tauri/src/managers/transcription.rs:
  • LoadedEngine::GigaAM now holds (GigaAMModel, VadChunked).
  • On GigaAM load, builds the pipeline SileroVad → SmoothedVad → VadChunked with parameters tuned for the longform CTC use-case.
  • On transcribe, calls chunker.transcribe(model, &audio) instead of model.transcribe(...) directly.
• Whisper, Parakeet, Moonshine, SenseVoice, Canary, Cohere — all unaffected.

[cjpais]

#1173