Issues in CSVToSampleLoaderVulnerability

Hello, There are some remaining issues with the CSVToSampleLoaderVulnerability. Mainly:

1. Currently raises an error as the reference_id parameter is accidentally fed twice when creating the DelayedSample (once explicitly, once as an element of the kwargs dictionnary).
2. (Minor) When loading reference samples, it still attaches a list of references to which the Sample should be compared, which doesn't really make sense. I think it is still fine to leave it that way as I guess this metadata won't be used anyway when running vanilla-biometrics.
3. Major one : when attack_type is None (cf. this snippet), we are supposed to compare the probe sample to every biometric reference. However, I believe this won't currently work as the list of all_references is built on the fly. So, when treating the first probes, this list of all_references could be almost empty. This should not pose problem if we call db.references() before ever calling db.probes(), but I think it is unsafe to rely on this assumption.

As I am planning to use this loader, I could work on fixing those aspects. 1) seems pretty straightforward to solve, however I am not sure how to approach 2) and 3) yet, so I am open to suggestions if you have some.

ping @ydayer