[benchmarks] Fix verifier error handling and OOM.

[ysiraichi]

In summary, this PR:

• Skips the actual experiment when running the verifier, i.e. goes directly into the verifier code, instead of running the experiment first (the experiment is still run in the end of the verifier, though)
• Handles errors when running the model on fp64, turning cosine similarity verification ON on any exception
• Renames a few confusing VerificationCode enum names

cc @miladm @JackCaoG @zpcore

[zpcore]

LGTM, thanks!