Use dense mass matrix when sampling

[bmorris3]

• pymc3 supports dense mass matrix adaptation thanks to @dfm's pull request.
• exoplanet has a handy exoplanet.get_dense_nuts_step function that you can use to get this functionality with minimal intervention.
• for background read on here and primarily here
• potentially we can also initialize the chains with the init="adapt_full" or "jitter+adapt_full" methods which should be related?

[bmorris3]

More thoughts:

• it's probably not a good API decision to wrap PyMC3 when we could just ask the user to use PyMC3 methods with minimal additional code. Consider ditching the Model.sample_<sampler> format in favor of a more PyMC3-ish idiom where we initialize with context managers explicitly on the user-facing API.
• based on @dfm's advice, we could choose to to avoid including exoplanet as a dependency by switching to pm.find_MAP() instead of Model.optimize. This would be the more pymc3-ish way. Need to check that this converges to the right solution as often as exoplanet.optimize.