jit compile x3 bootstrap types

[s3alfisc]

It should be possible to optimize this further - parts of the code can also be vectorized I believe.

Currently, I receive the following message:

NumbaPerformanceWarning: '@' is faster on contiguous arrays, called on (array(float64, 2d, A), array(float64, 1d, C))
  (inv_tXX_tXgXg[ixg,:,:] @ score_diff - delta_b_star)**2

[amichuda]

Actually, I've noticed significant speed improvements with np.einsum