Fix LLM streamProcess to emit tokens in real-time using callbackFlow

[sanchitmonga22]

Problem Statement

The LLM streaming implementation in LLMComponent.kt (line 512-529) buffers all tokens before emitting them, defeating the purpose of streaming.

Current behavior:

• Tokens are collected into a list via callback
• Only after streamGenerate returns are tokens emitted to the Flow
• Consumers wait for entire generation to complete before receiving any tokens

Expected behavior:

• Tokens should be emitted immediately as they arrive
• Consumers receive real-time updates during generation

Impact

• High - Core streaming functionality is broken
• Priority: P1

Proposed Solution

Restructure to use callbackFlow with trySend, similar to STTComponent.streamTranscribe (lines 342-382):

fun streamProcess(input: LLMInput): Flow<LLMGenerationChunk> = callbackFlow {
    ensureReady()
    val service = llmService ?: throw SDKError.ComponentNotReady("LLM service not available")
    input.validate()
    
    val options = input.options ?: RunAnywhereGenerationOptions(...)
    val prompt = buildPrompt(input.messages, input.systemPrompt ?: llmConfiguration.effectiveSystemPrompt)
    
    var chunkIndex = 0
    val sessionId = generateUUID()

    service.streamGenerate(prompt, options) { token ->
        trySend(LLMGenerationChunk(
            text = token,
            chunkIndex = chunkIndex++,
            sessionId = sessionId,
            isComplete = false
        ))
    }
    
    // Send final completion chunk
    trySend(LLMGenerationChunk(
        text = "",
        chunkIndex = chunkIndex,
        sessionId = sessionId,
        isComplete = true,
        finishReason = FinishReason.COMPLETE
    ))
    
    close()
    awaitClose { }
}

Implementation Plan

• Replace flow { } with callbackFlow { }
• Use trySend inside the callback instead of collecting to a list
• Emit final completion chunk with isComplete = true
• Ensure proper error handling with close(exception)
• Add awaitClose { } for proper cleanup
• Test streaming behavior in Android sample app

Files to Modify

• sdk/runanywhere-kotlin/src/commonMain/kotlin/com/runanywhere/sdk/components/llm/LLMComponent.kt

Related PR Comments

• [Kotlin-SDK] [Kotlin-Sample] Aligning Kotlin with iOS #233 (comment)

Success Criteria

• Tokens are emitted immediately as they arrive
• Chat UI shows progressive text generation
• No token buffering or delays
• Proper cancellation support