evaluation

[casperdcl]

• metrics
  • fix wall-time override (fixes metrics: exclude callback time #67)
  • compute per-iteration (follow-up to prepare for evaluation phase and run more data #91)
  • skip objective (fixes prepare for evaluation phase and run more data #91 <- fixes evaluation phase: skip objective #82)
  • skip normalised_change
  • add relative timing helper ("reset" scalar)
• raise StopIteration
  • increase timeout from 10 min to 1 hour
  • also stop if thresholds reached for >= 10 iteration window
• better logging ↔️ tqdm interaction
• move pass_index to petric for ease of use

[KrisThielemans]

seems fine.

I'd like to see some way to get metrics out of TB, but that might be for another PR (this is all about running it, saving, and stopping early enough)