Add opt-in Qwen3 ASR punctuated text output

[0xShug0]

Fixes #430

Summary

This adds an opt-in Qwen3 ASR request option, qwen3_asr.preserve_punctuation, for timestamped chunked ASR output.

By default, the existing --words-out behavior is unchanged: merged text_output is rebuilt from aligned word timestamps, so punctuation is omitted. When the new option is enabled, merged text_output keeps the punctuated ASR chunk text while still returning word timestamps.

Root cause

For multi-chunk Qwen3 ASR requests with timestamps enabled, the merge path rebuilds text_output from WordTimestamp.word. Those aligned words are word-only, so punctuation is lost even though the ASR chunk text has punctuation.

Behavior

Case	Command shape	Result
VAD + timestamps, default	--audio-chunk-mode vad --words-out	Unchanged: text_output is rebuilt from aligned words, so punctuation is omitted.
VAD + timestamps, opt-in	--audio-chunk-mode vad --words-out --request-option qwen3_asr.preserve_punctuation=true	text_output preserves ASR punctuation.
VAD without timestamps	--audio-chunk-mode vad	Unchanged: ASR punctuation is preserved.
Fixed chunking + timestamps, default	--audio-chunk-mode fixed --words-out	Unchanged: text_output is rebuilt from aligned words.

Validation

Built and validated with the Qwen3 ASR CLI path on CUDA using the repo sample audio asset.

Validation	Output
Default VAD + timestamps	Some call me nature Others call me Mother Nature I've been here for over four point five billion years Twentytwo thousand five hundred times longer than you
Opt-in VAD + timestamps	Some call me nature. Others call me Mother Nature. I've been here for over four point five billion years. Twenty-two thousand five hundred times longer than you.
VAD without timestamps	Some call me nature. Others call me Mother Nature. I've been here for over four point five billion years. Twenty-two thousand five hundred times longer than you.
Default fixed chunking + timestamps	Some call me nature others call me Mother Nature I've been here for over four point five billion years Twentytwo thousand five hundred times longer than you

Commands checked:

cmake --build build/debug -j$(nproc) --target audiocpp_cli
jq empty model_specs/qwen3_asr.json
audiocpp_cli --task asr --family qwen3_asr --backend cuda --audio assets/resources/sample_16k.wav --audio-chunk-mode vad --audio-chunk-seconds 5 --session-option qwen3_asr.forced_aligner_model_path=<Qwen3-ForcedAligner-GGUF> --text-out <out.txt> --words-out <words.json> --metrics --log --log-file <run.log>
audiocpp_cli --task asr --family qwen3_asr --backend cuda --audio assets/resources/sample_16k.wav --audio-chunk-mode vad --audio-chunk-seconds 5 --session-option qwen3_asr.forced_aligner_model_path=<Qwen3-ForcedAligner-GGUF> --request-option qwen3_asr.preserve_punctuation=true --text-out <out.txt> --words-out <words.json> --metrics --log --log-file <run.log>