Parallelizes LLM streaming

[Kuroki1931]

• Parallelizes LLM streaming requests during user speech.
• Starts new LLM generations from updated ASR context while keeping older in-flight generations
  alive briefly.
• Adopts the newest generation once it begins emitting tokens.
• Resets the oracle stream when switching to a newer generation.
• Drops stale generations so only the latest returned stream is fed to Moshi.
• Keeps Moshi/audio generation on the existing single-writer path to avoid concurrent LMGen access.

[yagumana]

I think this guard should reject older generations as well:

if gen_id <= self.adopted_gen:
    return

The caller checks gen_id > self.adopted_gen before calling _adopt_generation(), but that check happens outside _adopt_lock. If a newer generation is adopted while an older generation is waiting to enter this lock, the current equality-only check would still allow the older generation to set adopted_gen, causing adopted_gen to move backwards.

Since adopted_gen should be monotonic, this should be rechecked inside the lock.

[Kuroki1931]

Since this is a quick fix for the surface-level inference code, we will merge it for now. We will improve it later by integrating it with server_oracle.py.

[Kuroki1931]

This PR is for #4.