Add a filtering feature

[MichelDucartier]

Add the possibility to discard samples either:

• With a simple predicate: for instance, if a dataset has a column foo which contains a list of elements. We should be able to write something like predicate: {{ foo.len > 0 }} (or some kind of similar syntax)
• With a LLM that classifies as either "skip" or "keep" depending on some prompt