Resolve "Add caching of the `generate_posterior_sample_from_marginalized_likelihood` method"

This closes #569 (closed) and also fixes a number of other issues which popped up at the same time:

1. Closes #569 (closed) the reconstruction is now cached
2. The normalize_constraint_factor is not deterministic and changed between runs. This addresses this by calculating an average and rounding.
3. The check_prob and check_ln_prob methods where missing ratio factors for floats
4. Cleans up the cached_result behaviour: if the cache result exists, it is loaded and repeated steps are avoided
5. Remove the prior check from the cache file calculation: this always fails for marginalized runs because the prior gets dynamically updated. The behaviour now means that if a use changes the prior and re-runs, the cached file will be used. However, this should be fairly obvious to the user while the alternative: rerunning and slightly changing results is more problematic.
6. Simplify samples_to_posterior: move the check outside the function