Multiple improvements

[Kipok]

1. Handle decontamination by passing the "index" file with all problems only. This avoids a costly step of assigning labels to each solution, which can be many per problem as well as makes it easier to switch decontamination models/test sets.
2. Add require_num_files parameter to fill answer script to make it possible to use it in large pipelines and automatically fail if some of the random seeds from previous step are completely missing
3. Add more logging to fill answer script to be able to gain more visibility into what has changed after it ran
4. Fix some edge cases handling in llm judge script
5. Add reuse_code_exp parameter that allows to avoid uploading code to cluster and reuse it from another experiment. Requires latest nemo-run
6. Add mail_user/type to cluster configs to make it possible to get emails on failure
7. Some optimizations on the sft prepare default parameters + a couple of new filters
8. Make it possible to pass exclusive flag to run_cmd as it's often necessary when using cpu jobs (e.g. with prepare sft data on largest datasets we need as much memory/cpu as we can get)
9. Add WANDB_API_KEY as a default env var to pick up