Rework truncation based on log q

[mj-will]

Rework how the truncation based of the log-proposal probability (log q) to use the minium value of the training data.

Motivation

This could be an alternative to truncating the latent distribution. It should be less prone to over-constraining but may result in inefficient rejection sampling.

Changes

• Add methods for sampling from the latent distribution to FlowModel and BaseFlow
• Refactor FlowProposal.radius
• Add new keyword arguments to FlowProposal.backwards_pass
• Rename worst_q to min_log_q in FlowProposal.rejection_sample

Breaking changes

These changes are all to options that were not used by default and we generally recommended avoiding in the past, so I don't think they should impact users.

• This PR removes the current method for truncation that uses the log-probability of the worst point
• This PR removes the truncate keyword argument from FlowProposal
• This PR renames the worst_q keyword argument to min_log_q in FlowProposal.rejection_sample

To-Do

• Document the method
• Fix existing unit tests
• Add new unit tests
• Add a simple integration test

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (d5cd2b7) 93.26% compared to head (a1dcbb2) 93.29%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #338      +/-   ##
==========================================
+ Coverage   93.26%   93.29%   +0.02%     
==========================================
  Files          67       67              
  Lines        6300     6322      +22     
==========================================
+ Hits         5876     5898      +22     
  Misses        424      424              

Flag	Coverage Δ
unittests	93.29% <100.00%> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mj-will]

I'm deferring documenting this to a later PR since this would require a lot of additional documentation and the method is still not recommended.