Add type annotations to CUDA functions in straight guide and drop to 32-bit precision.

[mtbc]

Adds explicit rather than inferred type signatures to make it easier to experiment with 32-bit imprecision. With this in, all one need do for such an experiment is:

1. Change all the float64 to float32.
2. Before taking t2 and after taking t3, use neutron_array.astype(np.float32) and neutron_array.astype(np.float64) to convert the array of neutrons.
3. Check the types using something like,

   with open("/tmp/types", "w") as out:
   process_kernel.inspect_types(file=out)

[mtbc]

With the 32-bit version of the straight guide, for the tests' np.allclose I needed to downgrade the default atol by a couple of orders of magnitude. Presumably this gets even worse for a more numerically complex path.

[mtbc]

As one might expect, without the described changes, this PR doesn't much change execution speed nor register count.

[mtbc]

In experiments this morning, I'm finding that the described steps to drop to 32-bit slightly increase register count while improving speed by more than a third, so I'll add a commit here to do that and we'll see if the CI tests still pass.