Fix LongCat AudioDiT dtype, VAE, and attention API issues

[hlky]

Summary

Fixes the verified longcat_audio_dit review findings from #13580.

• use the transformer dtype for prompt embeddings, latents, latent conditioning, and normalized timesteps
• remove the internal torch.no_grad() from encode_prompt
• validate LongCat AudioDiT VAE downsampling_ratio against normalized strides and align defaults with [2, 4, 4, 8, 8]
• wire AudioDiTAttention / LongCatAudioDiTTransformer into standard diffusers attention processor APIs
• fix the slow real-checkpoint test to pass an instantiated tokenizer, not a tokenizer path
• add regression coverage for dtype plumbing, gradient-capable prompt encoding, VAE ratio validation, and attention APIs

Closes #13580.

Tests

• ruff check ...
• git diff --check
• make modified_only_fixup
• make fix-copies
• python -m pytest tests/pipelines/longcat_audio_dit/test_longcat_audio_dit.py tests/models/autoencoders/test_models_autoencoder_longcat_audio_dit.py tests/models/transformers/test_models_transformer_longcat_audio_dit.py -q
  • 67 passed, 17 skipped
• python -m pytest --collect-only tests/pipelines/longcat_audio_dit/test_longcat_audio_dit.py::LongCatAudioDiTPipelineSlowTests::test_longcat_audio_pipeline_from_pretrained_real_local_weights -q

The real-checkpoint slow test was collected but not run because it requires LONGCAT_AUDIO_DIT_MODEL_PATH, LONGCAT_AUDIO_DIT_TOKENIZER_PATH, and an accelerator.