More `RHO LOSS` flavors

[XianzheMa]

1. Make it possible to train the il model from last trigger's il model.
2. Enable sample-then-batch for RHO Loss.

Address https://github.com/eth-easl/modyn/issues/504

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 82.21%. Comparing base (6f7df09) to head (26df0f5).

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #510 +/- ## ========================================== + Coverage 82.19% 82.21% +0.01% ========================================== Files 214 214 Lines 9949 9956 +7 ========================================== + Hits 8178 8185 +7 Misses 1771 1771 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[github-actions[bot]]

^{( % to main) ( % to main)}