Distributed Tracing for remote scorers

[realark]

• Use real braintrust api client in TestHarness (pointed at VCR)
• do distributed traces for remote scorers
• create one span per scorer in evals

produces eval traces like this:

https://www.braintrust.dev/app/braintrustdata.com/p/andrew-misc/trace?object_type=experiment&object_id=ff8cb35c-4e3a-4eb4-ad45-16ce97a9a3b8&r=0b3378817b514de1c5367fb9ba07c60c&s=d5042e3790a0e674

[delner]

Personally not a big fan of explicit tracing code in the eval scorer implementation, but will defer on this (not blocking.)

[delner]

IMO, having tracing in the Eval scorer implementation seems like a bit of a code smell... it's creating hard coupling between Evals and OpenTracing which can be used separately from one another.

I'd recommend decoupling this through some appropriate abstraction; perhaps dependency injection, composition or some other pattern.

[realark]

I don't follow? Evals and otel already have a coupling in the sense that the data created by evals is captured mostly through otel traces, but in terms of what appears in the public apis for evals this doesn't change anything.

https://github.com/braintrustdata/braintrust-sdk-java/blob/main/src/main/java/dev/braintrust/eval/Scorer.java

I could add an additional method allowing explicit parent info to be passed when scoring:

// something like this
List<Score> score(TaskResult<INPUT, OUTPUT> taskResult, ParentInfo parentInfo);

But I'm not sure what that buys us? In this case it would make the surface area of the public api a bit larger and would only be invoked by callers within otel traces.

It would make sense in the context of a larger refactor to decouple otel from evals though. That seems beyond the scope of this PR so I'll press on, but let's chat about it some time