Streaming review

[arnabnandy7]

Part of review fixes for PR #1819

[igordayen]

@arnabnandy7 - thanks for addressing generate stream with thinking.
few comments to consider.
KDOC with examples could be helpful.
Performance test would be awesome, what is the avg overhead.
Thanks

[arnabnandy7]

@igordayen I've implemented all requested review changes:

• Expanded KDoc with chunked input → event output examples.
• Documented why extractAllThinkingBlocks() cannot be reused without whole-response buffering.
• Moved the parser into TaggedThinkingParser.kt.
• Added a deterministic 50,000-character scalability regression.
• Added warmed performance reporting without flaky timing assertions.

Verification:

Tests run: 49
Failures: 0
Errors: 0
BUILD SUCCESS

Performance across two warmed 30-run samples:

Input:             50,000 characters
Chunks:            1,001 × ~50 characters
Pass-through:      0.054–0.064 ms
Tagged parsing:    2.494–3.530 ms
Average overhead:  ~2.95 ms per stream
Per-chunk overhead: ~2.95 μs
Throughput:        ~16.6 million characters/second

[igordayen]

@arnabnandy7 - requesting changes reletated to documenting algorithm and reusability. the code is SO critical that readability should be addressed 100%.
Thanks for understanding.
Regards, Igor

[igordayen]

Will be working on a simple yet practical solution post 1.5.0 with full code reusability from the blocking thinking path.

Also, in the output, think is not triggered by native thinking.

Appreciate if you please keep dev on hold, thanks.