Further signal arraylias fix

[DanPuzzuoli]

Summary

First, a failing test for a perturbative solvers using JAX has been fixed by removing Array from the envelope function and instead using unp.

The rest of this PR fixes the remaining failures that all stem from the same issue (in both the pulse -> signal conversion code and in operator_collections.py). These were all the result of calling unp.append(x, y) with x being a numpy array and y being a JAX array. Due to unp dispatching on the first argument, this results in the evaluation of np.append(x, y), which fails when y is a JAX tracer.

The issue here is that what we fundamentally want here is multiple dispatch, but arraylias is single dispatch. To resolve this, I've added functions _preferred_lib and _numpy_multi_dispatch into dynamics/arraylias/alias that, given a list of arrays as arguments, will determine which array backend to use. Basically, for now it works as: if all arrays are of type "numpy" use numpy, but if any are of type "jax" then JAX is used. The current logic presupposes that these are the only two libraries. This will change later if more libraries are added.

[DanPuzzuoli]

Okay I think I got the docs to work as well. Sorry I think now this branch has conflicts with yours, though I suspect we made the same changes in the files with conflicts.