Add support for adapter switching and configurable encoders

[aviv1ron1]

What this adds

Speech input via a cascade, running as a single vLLM model — no separate speech server, no change to how the model is deployed or called. Audio is transcribed to text by an ASR encoder and the transcript tokens are spliced into the prompt before the decoder runs; the LLM only ever sees text.

The ASR encoder is distil-whisper/distil-small.en by default, but any HuggingFace automatic-speech-recognition pipeline model works via config alone (e.g. openai/whisper-small for multilingual) — no code change to swap it.

Capabilities

• Single-model cascade — audio in through the standard OpenAI/chat API (input_audio content parts); the chat template emits the <|audio|> marker and the processor replaces it with the real transcript tokens.
• Audio + adapters — audio requests route through the switch exactly like text, so an audio request can activate a LoRA adapter identically to its text equivalent.
• Config-carried ASR tuning — asr_pipeline_kwargs (pipeline construction, e.g. chunk_length_s) and asr_generate_kwargs (decode defaults, e.g. language/task) baked into config.json so the checkpoint is self-describing.
• Per-request language — a request may override decode kwargs via mm_processor_kwargs (allowlisted to language/task).
• Multiple audio clips per request — configurable ceiling (asr_max_audio_clips, default 32), each spliced at its own marker.
• Context-derived transcript budget — replaces the old fixed cap; per-clip budget = (max_model_len − generation_reserve − prompt) ÷ clips.
• Encoder-agnostic long-audio chunker — asr_self_chunks=false routes long audio through split → per-window transcribe → overlap-dedup merge, so a fixed-window backend inherits long-audio support (Whisper keeps its own internal chunking).
• Gated per checkpoint by asr_enabled — a non-audio checkpoint reports no audio modality and never loads ASR.

All of the above are settable at compose time (matching --asr-* flags) and editable directly in config.json.