Proposal: support Kyutai's REKI (Qwen3.5-2B + DiT expert with register memory) #45

Hi, after ACT I'd like to add REKI (paper, code). It's a Qwen3.5-2B VLM with a 0.2B flow-matching DiT expert, and it has a memory: 32 register tokens per frame in the expert, cached for the last 3 frames (1 s apart). The VLM keeps no state, only the registers persist.

Checkpoint: kyutai/reki-2b, plus LIBERO, RoboTwin and RoboCasa fine-tunes (BF16 safetensors). Weights are CC-BY-NC-SA 4.0 and the code is MIT, so the converter only reads weights users download.

Why: it's one of the few open VLAs with memory, and the released stack only runs on linux-64 with CUDA 12, so not on Jetson, Mac or CPU-only boards today.

Plan:

1. Qwen3.5 text stack (Gated DeltaNet + full attention, via ggml_gated_delta_net / ggml_ssm_conv, both in b11223) and its vision tower, as a module next to qwen3_lm.
2. The expert (AdaRMS with gated residuals, self- and cross-attention, SwiGLU) and the register path.
3. Converter from the released safetensors, vla_predict_check against the reference on reki-2b-libero over 3 consecutive frames, then a benchmark row.

One question before I start: the memory needs per-handle state. Would you accept vla_reset() in the C API (the register cache stored in the model handle, appended after each predict) and a reset flag on PredictRequest? Or would you prefer the caller to own the state (pass the cached register K/V in and get them back out)?