fix: speech to text live transcription

[IgorSwat]

Description

Various improvements & adjustments in Speech-to-Text module. The list of changes includes:

• Adjusting native implementation to the new format of Whisper models (single file, bundled encode & decode methods)
• Refactoring native implementation in order to support multiple STT models in the future
• Fixing an impropriate behavior of Whisper streaming

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

You can run the tests defined for Speech-to-Text module, as well as test it manually with the 'speech' demo app (SpeechToText screen).

Screenshots

Related issues

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

[msluszniak]

Some comments are not needed imo

[chmjkb]

Overall solid work, thanks 👏🏻
Left a couple of nits

[chmjkb]

food for thought: as we discussed a few days back, think about how we can make it work so that the native side doesn't need the model name, but accepts a bunch of configurable pipeline steps. no need to do this now IMO, but just a note.

Maybe we can have different ASR implementations based on whether the model does support timestamps or not?

[chmjkb]

I'm thinking whether we need to return std::vector from the encoder? Maybe we would just return a span. We wrap this in OwningArrayBuffer, which copies the data.

[IgorSwat]

I changed the encode & decode methods to return ExecuTorch Tensor - so no redundant copies now I guess.

[chmjkb]

Two more things:

1. I wasn't able to compile the app for Android (due to Norbert bumping minSdkVersion in RNET). You have to bump the minSdkVersion in the example app.
2. Once compiled, it doesn't ask for mic permissions (im using a Pixel 10) and silently fails.

[chmjkb]

Suggested change

	if (
	!modelSources ||
	!tokenizerSources ||
	!modelSources[0] ||
	!tokenizerSources[0]
	) {
	if (
	!modelSources?.[0] ||
	!tokenizerSources?.[0]
	) {

[chmjkb]

I think you should change the TS side as you're returnign a different thing from C++, for example:

  public async encode(waveform: Float32Array): Promise<Float32Array> {
    return new Float32Array(await this.nativeModule.encode(waveform));
  }            

Also, why did you switch back to type in SpeechToTextModelConfig?