Proposal: support LeRobot's ACT policy #37

ACT is probably the most common policy on SO-100/SO-101 arms, and it isn't in vla.cpp yet. I'd like to add it, but since it isn't a VLA (no language input), I wanted to check it's in scope first.

Why it could be worth it: PyTorch ACT is already fast on a Jetson (~59 ms per 100-action chunk on my Orin NX), but it still needs a CUDA PyTorch build for the board. On CPU-only boards like the one in #31, vla.cpp would be the lighter option.

What I have in mind, following "Adding an architecture" in CONTRIBUTING:

• scripts/convert_act_to_gguf.py for LeRobot ACT checkpoints: ResNet18 with its batch norms folded into the convs, the normalizer stats from the processor files, and no VAE encoder (it's training-only, so the latent is zero at inference). That leaves ~34M of the 52M parameters.
• src/models/act.cpp: ResNet18 on the conv path Octo uses, the 4-layer encoder, and the 1-layer decoder with 100 learned queries, similar to TurboVLA's ACT decoder but post-norm. No flow-matching steps.
• Checked against PyTorch with vla_predict_check, then on my SO-101. I have an ACT checkpoint at 9/10 on the same task as the SmolVLA run in Deployment pipeline for SmolVLA on a real SO-101 arm (Qualcomm Dragonwing IQ-9075, ARM CPU) #31.

Would you take this? And if so, should --arch act in the LeRobot client fork come with it?