fix: treat lfm2 as ambiguous model type (LLM + embedding + audio)

[unstoppablesssss]

Problem

LFM2.5 Instruct models (e.g. LFM2.5-1.2B-Instruct-MLX-4bit) are misdetected as embedding models and fail on /v1/chat/completions requests.

Root cause: lfm2 is listed in EMBEDDING_MODEL_TYPES, so any model with model_type: lfm2 is unconditionally classified as embedding — even when the architecture is Lfm2ForCausalLM (a standard LLM).

LFM2 family variants

Model	model_type	Architecture	Expected type
LFM2.5-1.2B-Instruct	lfm2	Lfm2ForCausalLM	llm
LFM2.5-1.2B-Thinking	lfm2	Lfm2ForCausalLM	llm
LFM2 embedding	lfm2	BertModel etc.	embedding
LFM2 audio (STS)	lfm_audio	LFM2AudioModel	audio_sts

Fix

1. Move lfm2 from EMBEDDING_MODEL_TYPES → AMBIGUOUS_EMBEDDING_MODEL_TYPES (same pattern as qwen3, gemma3-text)
2. Narrow the LFM2 audio catch-all from startswith(lfm) to exact match lfm_audio to prevent LLM variants from being misclassified as audio_sts

Verification

Tested locally on oMLX 0.3.7 with LFM2.5-1.2B-Instruct-MLX-4bit:

• Before: Discovered model: LFM2.5-1.2B-Instruct-MLX-4bit (type: embedding, engine: embedding)
• After: Discovered model: LFM2.5-1.2B-Instruct-MLX-4bit (type: llm, engine: batched)

Chat completions now work correctly.