Parallelize hessian estimation

[mattwigway]

Hessian estimation is the slowest part of the model fitting process, and it's single-threaded. This seems like something where parallelization should be possible. Investigate.

[mattwigway]

Parallelization is definitely possible with the new linear_hessian code introduced in #52. I think the right way to do it is to pre-calculate all steps and store them in shared memory, along with the utility, and then use a multiprocessing pool to chug through all the hessian results in parallel. The challenge is how to do this without having multiprocessing copy the whole MNLFullASC object.

[mattwigway]

Since Numpy is good about releasing the GIL, I can use threads and not have to deal with shared memory.