n_check_point highly variable

The default n_check_point calculation is highly variable. It's based on some factor times the log_likelihood eval. This means that between two runs n_check_point will almost always be different. This causes the cache check to always fail.

We could fix this in two ways

• Remove n_check_point from the cache check (i.e. if you have change the check pointing it won't rerun)
• Make the n_check_point more stable. I.e., round log_likelihood_eval to the nearest order of magnitude before setting n_check_point. This also has the advantage the n_check_point will be a sensible number (i.e. 10000 rather than a random integer).

My vote is we do both. Any other thoughts? @colm.talbot @paul-lasky @moritz.huebner