Enable parallel sampling of chains

[TobyBoyne]

Currently, we do not sample chains in parallel. This represents a huge wasted opportunity for speeding up sampling. However, initial efforts haven't been able to work. See the fix-parallel-sampling branch.

Running the sampling raises the error:
No definition for lowering UniTuple(int64 x 4).shape.

[TobyBoyne]

Waiting on numba/numba#9728, or converting to using @jitclass

[TobyBoyne]

Parallel compilation in numba is still quite temperamental. This has been disabled for now. We note that, for 100 points, taking 500 samples takes about 1.5-2 seconds. For 8 parallel chains, sampling therefore takes <20 seconds. This is an acceptable time taken (for now), as training across 100 iterations should take less than 30 minutes.

This also means that training is of a similar time scale to the optimization.