Lots of type instability for `Float32` arguments

for one of the more complicated examples, most of the branches of `besselh(0, 1, 1f0)` are type unstable since they return a random mix of `Float32` and `Float64`. Ideally Float32 inputs would do the computations in Float32 as much as possible (and use the lower orders allowed by the reduced precision requirements.

This is especially important for GPU use cases where Float32 is more abundant.