Refactor vectorised HMC implementations using `AcceleratedKernels.jl`

[yebai]

I recently discovered the AcceleratedKernels package, which provides a unified interface for parallelisation on CPUs, clusters, and GPUs. We could consider switching to AK.map! and other AK high-order functions for the vectorised HMC implementation.

Related discussions: #390 (comment)

Question: would AK.map! utilise SIMD similarly to the current AHMC vectorisation?

[yebai]

Also related: https://github.com/tkf/ThreadsX.jl