Feature/analytical mean

[edaub]

Implements analytical mean functions (and a few other things needed for this). Addresses #96.

New in this PR:

• Mean functions are fit analytically.
• Mean function priors (as multivariate normal distributions) are now implemented. The analytical mean solution depends on manipulating the matrices associated with these.
• Some additional calculations are abstracted into functions in the linalg submodule. This should clean up some of the GP log posterior computations and make the logic easier to understand.
• Remove fit_cov option from GPParams and GPPriors.
• Docstring improvements in a number of classes.

[edaub]

@nbarlowATI So it turns out that with a nugget there is no longer an analytical solution to the covariance (something I realized only when my tests kept failing because the analytical and numerical solutions never agreed!). Some of my previous work added some additional functionality to allow for this, which I should probably just drop for the sake of simplicity. If you haven't reviewed this yet, give me a chance to clean up that part of the code.

[edaub]

@nbarlowATI Should be fixed now!

[nbarlowATI]

Looks good!
Great that the docstrings are now pretty much complete, and the refactoring all seems sensible. All tests pass for me too.
Only suggestion I'd make is to remove, or redefine n_params, so that it is no longer misleading (maybe just have it as n_data + n_mean ?)