What does this PR do ?

Unify ModelPT.validation_step_outputs (and test_step_outputs) to always return a list of lists, so a single dataloader is simply the N=1 case and subclasses no longer need to branch on the output shape. Normalize both _validation_dl and _test_dl to Optional[List[DataLoader]] via their respective resolvers.

Collection: Core, ASR, TTS, Audio

Changelog

• modelPT.py: validation_step_outputs / test_step_outputs properties always return [[] for _ in range(num_dl)]; on_validation_epoch_end / on_test_epoch_end use len() == 1 instead of isinstance(..., dict) for single-vs-multi dispatch; empty-output guard updated to all(len(o) == 0 for o in ...) since [[]] is truthy; empty dataloader buckets skipped in multi-DL loop
• model_utils.py: resolve_validation_dataloaders and resolve_test_dataloaders wrap bare DataLoader into [DataLoader] at both single-value paths, normalizing _validation_dl and _test_dl to Optional[List[DataLoader]]
• modelPT.py (setup_multiple_validation_data): type annotation updated; isinstance guard simplified to truthiness check after normalization
• 15 model files (ASR, TTS G2P, Audio): remove if/else branching in validation_step / test_step; always use self.validation_step_outputs[dataloader_idx].append(...)
• transformer_bpe_models.py: remove isinstance(outputs[0], dict) normalization loop in multi_validation_epoch_end — base class now iterates dataloaders and calls it once per DL
• audio_to_audio.py: simplify _get_num_dataloaders (both val and test) and logging callback setup after normalization
• fastpitch.py, magpietts.py, magpietts_preference_optimization.py: add RuntimeError guard for len(validation_step_outputs) != 1; add early-return on empty outputs; use self.validation_step_outputs[0] consistently
• ssl_models.py: fix EncDecMaskedTokenPredModel.test_step — was appending to validation_step_outputs instead of test_step_outputs
• Test models (test_ema.py, check_for_ranks.py, test_ptl_stateless_timer.py): override multi_validation_epoch_end instead of on_validation_epoch_end; base class handles iteration, clearing, and per-DL prefix

Usage

No API changes for single-dataloader models — dataloader_idx=0 is the default. Subclasses should use the [dataloader_idx] indexing pattern:

# In validation_step:
def validation_step(self, batch, batch_idx, dataloader_idx=0):
    metrics = self.compute_metrics(batch)
    self.validation_step_outputs[dataloader_idx].append(metrics)
    return metrics

# In multi_validation_epoch_end (preferred over on_validation_epoch_end):
def multi_validation_epoch_end(self, outputs, dataloader_idx=0):
    avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
    self.log("val_loss", avg_loss)