Remove PubMedQA built-in benchmark code

[arschles]

It already exists in the registry server, and should be executed from there rather than the code.

We should make sure that the following are fixed first:

• On qt resume, ensure provenance and reproducibility for remote evals #63
• Add a label scoring style to custom_nocode benchmarks #59

And, after #59 is fixed, the remote benchmark definition should use the new label style rather than the existing exact_match style. Doing this will improve the percentage of parsed responses and thus improve benchmark performance.