PUMA / CONIC semantic

[titusgriebel]

@anwai98
For CONIC:

• there is a pre-defined split that was generated with a specific logic. It creates 10 folds whereas the IDX 0 was used to train the HoVerNet baseline of the challenge (I would suggest sticking to it for our training and evaluation). This is implemented in a separate patho-sam branch. I was unsure whether to add this to the torch-em script.
• the download logic currently does not take into account that download could be set to false without data being there already. Maybe you have an elegant way to implement this.

[titusgriebel]

@anwai98 I updated the scripts as discussed!

[anwai98]

Hi @titusgriebel,

Thanks for the PR. I will check it out in a bit. Meanwhile, would be nice to set up the downstream script for experiments already, so that once this is finalized and merged to main, we can fire up the GPUs! ;)