feat: Add CLAP (Contrastive Language-Audio Pretraining) support

[JeniaJitsev]

Summary

Add audio-text contrastive learning (CLAP) to open_clip, extending it beyond vision-text to support audio-text modality pairs.

Audio encoders

• HTSAT (tiny/base/large) with attentional feature fusion (DAF/AFF/iAFF)
• Whisper (tiny/base/small/medium/large) with optional pretrained OpenAI weight loading

Text encoder

• RoBERTa-base via HuggingFace transformers (mean_pooler / cls_pooler)

Training pipeline

• WebDataset audio pipeline (--dataset-type webdataset-audio)
• Synthetic audio dataset for CI/smoke tests (--dataset-type synthetic-audio)
• Configurable preprocessing: --data-truncating {rand_trunc,trunc,fusion}, --data-filling {pad,repeat,repeatpad}
• Fusion mode for longer audio clips (--enable-fusion --fusion-type {daf_2d,aff_2d,iaff_2d,...})
• Pretrained audio encoder loading (--pretrained-audio)
• Audio-aware profiling

Model configs

• 13 HTSAT/CLAP/Whisper JSON model configurations
• CLAP model class (mirrors CustomTextCLIP, with audio tower instead of vision)

Key design decisions

• CLAP class returns image_features key for ClipLoss compatibility — zero changes to loss/training pipeline
• AudioTower wraps encoders (HTSAT or Whisper) with a 2-layer MLP projection
• Audio models are automatically excluded from vision test suites
• SyntheticAudioDataset generates random waveforms for CI without audio dependencies

Test plan

• test_training_clap passes (HTSAT-tiny, synthetic-audio, 1 epoch, CPU)
• Whisper-tiny smoke test passes (synthetic-audio, 1 epoch, CPU)
• Pretrained Whisper weight loading verified (67/70 keys for tiny, 187/190 for small)
• Whisper layer counts verified against official OpenAI spec (4/6/12/24/32)
• Existing vision tests unaffected (audio models filtered out)
• Multi-GPU distributed training (tested on NVIDIA GH200 cluster, up to 1024 GPUs)
• Evaluation on UrbanSound8K, AudioCaps, Clotho benchmarks (via CLIP Benchmark custom fork)

[JeniaJitsev]

Using implementations from https://github.com/marianna13/open_clip/tree/clap . @rwightman @rom1504 @marianna13