Initial performance optimizations for frame selection pilots

First attempts at improving the abysmal performance observed while applying frame selection at the proposal stage.

Profiling based on py-spy and !7 (merged) pointed to two major culprits:

• Reading the full list of keys of AGIPD sources

  • Fixed in 0509da5e with an optimized version of SourceData._get_index_group_sample (@kluyvert I will upstream a general verison of this in EXtra-data later.)
  • Fixed in 6d0fd2bd by skipping an over-eager check for compliance

• Reading very large datasets (e.g. image.data)

  • Fixed in 20e6d23b through parallel reading of large datasets (> 1 GiB for now) with pasha

Both measures sped up processing of large runs by ~3x.