ENH: implement JAX modified Bessel function via exponentially modified version

[fonnesbeck]

The modified Bessel function of the first kind is not currently available in Jax, but there is an exponentially-scaled version such that:

ive(v, z) = iv(v, z) * exp(-abs(z.real))

However ive is not available in PyTensor. Implementing it would allow us to have iv from this transformation.

Context for the issue:

The Bessel functions are useful in places like the Hilbert space approximation kernels for GPs. It would be valuable to be able to sample these with the Jax backend for performance reasons.

[ricardoV94]

There are two things. We can easily add an ive helper to pytensor.tensor without implementing new Ops

Separately from that, if you want to use iv in JAX we can use the (tfp) ive in the JAX dispatch of iv. Only issue is we need to add tensorflow probability as an (optional) dependency? Also I haven't tried to use it with JAX, so I don't know exactly how the two can me married.

I believe @ColCarroll might have some pointers?

[fonnesbeck]

Another big one we are missing is the error function erfc which keeps us from using the normal CDF in any JAX sampling. There's no JAX implementation either, so we could go the TF route there too.

[twiecki]

Couldn't we copy it over from tfp?