Fix: restore token-by-token streaming for chains using createTextOnlyOutputParser

[HenryHengZJ]

• createTextOnlyOutputParser() used a RunnableLambda which accumulates all LLM tokens before yielding, breaking progressive streaming in any chain that relies on streamLog() (e.g. Conversational Retrieval QA Chain)
• Replaced with a proper Runnable subclass (TextOnlyOutputParser) that implements _transform() to yield each token as it arrives
• Reasoning-model filtering (stripping thinking/reasoning blocks) is preserved

[gemini-code-assist]

Code Review

This pull request refactors the createTextOnlyOutputParser utility to improve streaming support by replacing the RunnableLambda implementation with a dedicated TextOnlyOutputParser class. The logic for extracting text from message chunks has been moved to a helper function. A review comment suggests enhancing type safety by replacing the any type with a type guard during the filtering of content blocks.

[HenryHengZJ]

This is a fix coming from here

[estebanjosse]

I tested this PR locally and can confirm it resolves #6070 on my side.

Streaming is now working again as expected in the Conversational Retrieval QA Chain (token-by-token output in the UI).

For context, I initially investigated this issue and opened an alternative fix in #6089, but I only saw this PR afterwards.

This approach also resolves the issue correctly 👍

I also added a test harness covering this behavior in my PR — I’d be happy to contribute/adapt those tests to this PR if useful.