Streamline VCF annotator ABC definition

In #502 I defined an abstract VCF annotation interface where implementations could produce side-effects on a per-variant basis as well as upon the entire collection of variants. Of course, to do the latter, you need to maintain a collection of all the variants for the duration of annotation. Since no side effects are defined on the base class, I defined a class variable that needs to be switched on to enable collection-based operations, in addition to defining the hook function itself.

This of course becomes a problem for large VCFs, because Pydantic (especially v2) objects occupy a lot of memory. And it's a problem that this value is defined at a class level even though inputs are given on a per-annotation basis, so you might not even want to be doing the collection. In #545 I added an additional, optionally-implementable method for implementations to conditionally suppress that collection. I think this interface/related operations could be improved in a future major release (+ some smaller possible improvements in the meantime)

* Check for collection suppression on a per-annotation basis. Ditch the class variable, just redefine the `should_collect_alleles` method.
* The allele collection should contain serialized objects, rather than the full Pydantic object.
* NDJSON should be written on a per-variation basis rather than being added at the end.
* Maybe don't support dumping as a PKL (it's limited to just re-loading the Python object, it theoretically raises security issues). Maybe we shouldn't be supporting this whole "perform on a collection" hook at all, insofar as non-small VCFs are going to be expensive to hold variants in memory for. Maybe it's worth reorienting our output framework on things that can be streamed to.