feat: Add Langfuse Integration and Fix Audio Processing

[MrOrz]

This PR adds Langfuse integration and includes several fixes for audio processing and OpenAI-related functionality.

Main Changes

Langfuse Integration

• Initial Langfuse setup. However, this requires us to upgrade OpenAI SDK
• Use AI response ID as trace ID, so that we can score these traces easily in the future

OpenAI & Testing Updates

Upgraded OpenAI SDK, which broke a few things and need repairs.

• Fixed OpenAI SDK method typo
• Updated OpenAI mock implementations to reflect API changes
• Fixed circular reference and formatting in CreateAIReply.js mocks
• Temporarily skipped audio transcription test due to fluent-ffmpeg issue

Build & Configuration

• Enabled experimental-vm-modules for Jest compatibility with Langfuse SDK
• Fixed Langfuse import issues

The changes primarily focus on improving audio processing reliability and integrating Langfuse while maintaining test coverage and fixing various related issues.

Example Langfuse traces

• Whisper transcript: https://langfuse.cofacts.tw/project/cm3e6a2190001fdga2ruendgd/traces/D4iRcJQBpiYdBld_hVwQ
• AI reply: https://langfuse.cofacts.tw/project/cm3e6a2190001fdga2ruendgd/traces/EYiRcJQBpiYdBld_ylzv

[coveralls]

coverage: 82.699% (-0.7%) from 83.364%
when pulling 61a3497 on langfuse
into 5dc1e5e on master.

[MrOrz]

After deploying to staging and testing it using videos, I've got some errors. Staging bot is not working now. Going back to draft state.

Issue #1: output stream closed

Similar to fluent-ffmpeg/node-fluent-ffmpeg#1135 , somehow Whisper will close connection early and triggers error on fluent-ffmpeg.

It seems that the logic still works though, may consider ignoring the issue.

Issue #2: AI reply bug

Seems that we should fix CreateAIReply. The LLM call is mocked so that the change to the interface is not captured in the unit test.