Update trainset sampling

[nadahlberg]

This PR makes several changes to how the trainset is sampled:

1. Now each sampling method tracks it's own status, so that we don't need to resample anytime anything changes. So sampling from an annotation decision or a querystring sampler only happens once.
2. Heuristic bucket sampling is not idempotent. This is also needed so that we can preserve the trainset across updates.
3. Querystring samplers: User's can now anchor on a querystring search and save it as a "sampler" so that at least N examples that satisfy the query will be added to the trainset.
4. This PR also replaces the DSPy predictor with a simpler few-shot annotation agent.

[nadahlberg]

Fixes: #142 #141 #140