Have more versions of `RHO-LOSS`

[XianzheMa]

1. We can enable sample-then-batch for RHO-LOSS.
2. When we select points, we can select depending on the RHO Loss as weight, rather than just select the top k.
3. We can finetune the previous IL model instead of training a randomly initialized one

Proposed from https://github.com/eth-easl/modyn/pull/491#discussion_r1637887297