Numba-ify the log likelihood function

This uses the numba JIT package in an attempt to speed up the core of the log_likelihood function (and also the Bayesian blocks algorithm).

In some tests on fake Gaussian noise, which does not get split into any chunks the numba version is ~10% slower than using numba for calculating the likelihood. For the full O1 data set, which gets split into many chunks, using numba gives a speed up of over 4 times!

This closes #9 (closed) by using numba instead of Cython.