Implement sampling in float32

[aseyboldt]

Right now everything in nutpie (the sampler and the logp functions) are working in float64. For some models we could get decent speedups if we also support float32. Pytensor can already compile float32 logp functions, but we have to change the python wrapper to make use of that. We also have to add a <F: Float> type parameter to some operations in nuts-rs, and figure out how to do the simd functions with that.