fix: persistent_workers crash at num_workers=0 + warn on silent token truncation

[rohitsynth1208]

Is this PR adding new feature or fix a BUG?

Fix BUG (two small, independent fixes in fish_speech/datasets/semantic.py).

Is this pull request related to any issue? If yes, please link the issue.

Not filed as a separate issue, but related in spirit to #1163 (pretrained weights being silently destroyed) - this PR is about a different pair of silent failure modes in the dataset/dataloader path that we hit while LoRA fine-tuning text2semantic on a small (~1,000 clip) custom dataset.

1. persistent_workers=True hardcoded, crashes at num_workers=0

SemanticDataModule.train_dataloader()/val_dataloader() hardcode persistent_workers=True regardless of num_workers. PyTorch's DataLoader raises immediately if persistent_workers=True is passed with num_workers=0:

ValueError: persistent_workers option needs num_workers > 0

num_workers=0 is a legitimate, common choice for small datasets (few shards) or when debugging - the finetune docs' own recipe doesn't forbid it. Made persistent_workers conditional on num_workers > 0, matching the flag's actual precondition.

2. Silent truncation with no warning in TextDataCollator.batchify()

_tokens = example[tokens_key][:, :max_tokens_length]

This slices any example longer than max_length with zero log/warning. Since a packed example's token sequence includes the full audio-codec token stream for its sentence(s) - not just text - and AutoTextSemanticInstructionIterableDataset.sample_data()'s num_samples = self.max_length // 20 heuristic only estimates length from text tokens (# estimate that each sample is at least 20 tokens), it's easy to end up with examples that exceed max_length once audio tokens are counted, and get silently truncated mid-audio - corrupting the training target with no indication anything went wrong.

Added a log.warning(...) when this truncation actually fires, so users get a clear signal to raise max_length (or reduce num_samples) instead of unknowingly training on truncated audio. Purely additive/observability - no change to actual training behavior.

Testing

Both fixes verified against a real fine-tuning run (LoRA, text2semantic_finetune, ~1,000 clip custom dataset, num_workers=0, H100): the first fix is required just to get training running at num_workers=0 at all; the second warning was verified to correctly fire on examples we confirmed (by manual inspection) were being truncated mid-audio under the default heuristic.

— Submitted by SynthioLabs

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}