feat(firebase-ai): add transcriptions to Live API sample and reduce abstraction

[thatfiredev]

There are 2 main changes in this PR:

• Add support for showing transcriptions on screen for the Gemini Live API samples.
• Delete the BidiViewModel class to reduce abstraction. This should make the samples more copy-pastable.

[gemini-code-assist]

Code Review

This pull request refactors the live streaming feature by replacing the abstract BidiViewModel with two dedicated ViewModels, StreamAudioViewModel and StreamVideoViewModel, and adding support for real-time transcriptions in both audio and video screens. The review feedback highlights several critical improvements: managing the conversation lifecycle with DisposableEffect and a lifecycle-aware CoroutineScope to prevent leaks, avoiding blocking the main thread with runBlocking during ViewModel initialization, and replacing animateScrollToItem with scrollToItem to prevent UI stuttering during rapid transcription updates.

[gemini-code-assist]

Declaring liveSession as a non-nullable property initialized via runBlocking in the init block blocks the main thread during ViewModel creation to establish a network connection. This can cause UI freezes or ANR (Application Not Responding) errors. Instead, declare liveSession as lateinit var (or nullable) and connect to the live model asynchronously using viewModelScope.launch.

Suggested change

	private var liveSession: LiveSession
	private lateinit var liveSession: LiveSession

[gemini-code-assist]

Declaring liveSession as a non-nullable property initialized via runBlocking in the init block blocks the main thread during ViewModel creation to establish a network connection. This can cause UI freezes or ANR (Application Not Responding) errors. Instead, declare liveSession as lateinit var (or nullable) and connect to the live model asynchronously using viewModelScope.launch.

Suggested change

	private var liveSession: LiveSession
	private lateinit var liveSession: LiveSession

[gemini-code-assist]

Using animateScrollToItem for rapidly updating streaming text (like live transcriptions) can cause severe UI stuttering, lag, and high CPU usage because smooth scroll animations are repeatedly interrupted and restarted. Using scrollToItem is much more efficient and ensures smooth rendering during active streaming.

Suggested change

	LaunchedEffect(transcriptions.size, transcriptions.lastOrNull()?.text) {
	if (transcriptions.isNotEmpty()) {
	listState.animateScrollToItem(transcriptions.size - 1)
	}
	}
	LaunchedEffect(transcriptions.size, transcriptions.lastOrNull()?.text) {
	if (transcriptions.isNotEmpty()) {
	listState.scrollToItem(transcriptions.size - 1)
	}
	}

[gemini-code-assist]

Using animateScrollToItem for rapidly updating streaming text (like live transcriptions) can cause severe UI stuttering, lag, and high CPU usage because smooth scroll animations are repeatedly interrupted and restarted. Using scrollToItem is much more efficient and ensures smooth rendering during active streaming.

Suggested change

	LaunchedEffect(transcriptions.size, transcriptions.lastOrNull()?.text) {
	if (transcriptions.isNotEmpty()) {
	listState.animateScrollToItem(transcriptions.size - 1)
	}
	}
	LaunchedEffect(transcriptions.size, transcriptions.lastOrNull()?.text) {
	if (transcriptions.isNotEmpty()) {
	listState.scrollToItem(transcriptions.size - 1)
	}
	}