fix(higgs_audio_tts): share one sampling seed across chunks of a request

[mwzkhalil]

What this fixes

Reported in #471: Higgs TTS voice cloning "switches to a different voice" partway through longer text, and separately, generation can hit Higgs TTS generation reached max_tokens before EOC with no indication of how to fix it.

Root cause of the voice drift: HiggsTTSSession::run() splits long text into chunks (chunk_text_request) and calls generator_->generate() once per chunk. When a caller does not pass --seed, each of those calls independently resolves its own random seed:

sampling.seed = request.options.seed.value_or(runtime::random_u64_seed());

So an unseeded multi-chunk request samples every chunk against the same reference-voice conditioning but with unrelated entropy each time. That's consistent with what's reported: it gets worse with a smaller --text-chunk-size (more chunk boundaries = more independent re-rolls).

The fix

HiggsTTSSession::run() now resolves one seed up front when the request didn't supply one, and copies it onto every chunk's options before generation — the same as what already happens when a caller passes --seed explicitly (chunks already shared it via the copied options map in that case). This only changes behavior for the previously-undefined unseeded path; explicit --seed behavior is bit-for-bit unchanged.

I don't have a way to run the actual 4B GGUF weights in this environment to confirm the audio-level fix end to end, so I'd appreciate it if someone who can reproduce #471 tries this branch against the same Arabic long-form + voice-ref case. The mechanism is sound (verified by reading generator.cpp's seed resolution and session.cpp's per-chunk request construction), but I want to flag that the perceptual result is unverified on my end.

Also included

The max_tokens / EOC failure now names the configured value and points at --max-tokens / --text-chunk-size (or their server request-option equivalents) instead of leaving the caller to guess:

Higgs TTS generation reached max_tokens (2048) before EOC for this text chunk; raise it
with --max-tokens on the CLI or the "max_tokens" request option on the server, or lower
--text-chunk-size / "text_chunk_size" so each chunk needs fewer generated frames

This is a message-only change (same failure condition, no behavior change) — it was previously easy to hit with default settings on dense scripts (Arabic in the report) with no clue what to adjust.

Testing

• CI (Linux, macOS, Windows, Nix) green on this branch: https://github.com/mwzkhalil/audio.cpp/actions?query=branch%3Aci%2Fhiggs-tts-chunk-seed-consistency
• No existing test asserts exact output for an unseeded multi-chunk request (the one warm-bench test that checks seed always forwards an explicit value), so this doesn't touch any golden-output comparison.
• I could not run the actual model (no GGUF weights / GPU in this environment) — see the callout above.

Ref #471

[0xShug0]

@mwzkhalil Unless I’m missing something, setting the seed won’t fix this issue. Please provide wavs generated by higgs tts on CPU with your PR applied using the test case provided in the issue.