Actually fixing reverse AD

[cgiovanetti]

Fixing #31. The VeryChord thing was a red herring, turns out the problem was with the inf padding in the hyrex outputs (which is why we were only seeing NaNs in a couple of densities that touch e.g. xe).

Here's what was happening:

1. On the backward pass of reverse AD, the RHS of the perturbations equations get evaluated again. This already opens us up to the double-where NaN issue in JAX
2. Sometimes, t1 in evolution_one_k ends up pushing these arrays very close to lastval, getting closer and closer for larger k. lensing=True adds a few additional larger k. This increases the risk of fast_interp needing to interpolate at an index beyond lastval.
3. If we are unlucky with our grid so that we do end up beyond the lastval boundary, fast_interp gives an inf. This doesn't show up on forward passes since this isn't the branch of the jnp.where that actually gets evaluated, but because we're taking a gradient JAX freaks out unless both branches of where are finite.

So the fix was pretty simple, once we have the outputs from HyRex I just patch lastnum on to all of the inf values of the array so this doesn't trigger. No change to speed or accuracy, gives finite gradients.

Also got rid of the previous VeryChord tolerance fix, it doesn't actually get us anything.