feat: add C2LLM embedding model (codefuse-ai/C2LLM-7B)

[StartupBros]

Adds support for codefuse-ai/C2LLM-7B (model_type: c2llm), currently tied for #1 on the MTEB Code leaderboard.

What it is

A Qwen2 decoder backbone with a PMA (Pooling by Multihead Attention) head instead of mean/last-token pooling. A single learned seed vector attends over the token states, so the pooled vector is a learned query over the sequence rather than a positional pick.

The backbone is Qwen2, not Qwen3 — q/k/v carry biases and there are no per-head q_norm/k_norm tensors.

Parity details worth reviewing

Five things in the reference are load-bearing and none is the obvious choice. Each is documented in the module docstring:

• the attention scale is sqrt(dim_V) — the full compressed width (3584), not the per-head width (112)
• the softmax is re-masked afterwards (A = A * pad_mask), so rows over padded input deliberately do not sum to 1
• the residual adds the projected seed, not the raw seed. The model instantiates PMA_v2/MAB_POST_v2; the PMA/MAB_POST classes directly above them in the same upstream file add the raw seed instead
• the output feed-forward is ReLU, unlike the SiLU backbone
• pma_norm is false, so the reference returns an unnormalized vector. text_embeds is normalized to match the library contract; pooler_output keeps the raw vector

Verification

Against the real checkpoint on an M3 Ultra. The upstream reference cannot be imported for an end-to-end comparison — modeling_c2llm.py declares deepspeed and peft imports that transformers' check_imports enforces — so each half was compared separately, which also localizes any failure:

half	reference	result
PMA head	verbatim transcription of MAB_POST_v2	cosine 1.0 on both rows, incl. a left-padded one
backbone	transformers' own Qwen2Model, same tensors	min cosine 1.0 over 41 unpadded positions; 0 missing / 0 unexpected on load

Parameter count is 7,122,031,616 — the checkpoint's 7,667,028,992 minus the 152064x3584 lm_head that sanitize drops.

Also serving-verified end to end: loaded under oMLX 0.6.3 and returned a 3584-dimensional vector over /v1/embeddings.