feat: add CAPI imagenet benchmark (#1998)

* feat: add CAPI imagenet benchmark

* docs: note the encoder RoPE deviation in the CAPI benchmark

The encoder is a standard masked ViT; the reference also applies RoPE inside the encoder. Matches the deviation already documented for the CAPI examples (#1996).

* refactor: use vit-small encoder in CAPI benchmark

Align the encoder with the other vitb16 benchmarks (dino/dinov2/ibot) and size the predictor to it: num_heads=6 and depth=6 (half the encoder depth), following CAPI section 4.1. Addresses review.