Optimise missing sources

[mauch]

Modify get_src_skyregion_merged_df to improve speed and memory efficiency. This also preps this function for furture checkpoint and restart operations.

Changes are:

• Replace source/image/epoch strings with smaller int32 indices for the duration of the function and use int64 for datetime sorting rather than a Timestamp object
• Use numpy arrays and vectorised lookups and tricks inplace where possible to avoid large pandas dataframe copies during merge/groupby/sort operations
• Convert the output missing sources dataframe to dask before exploding the image_diff column and returning source IDs to srtings at the end of the function
• Get rid of the two explodes on missing_sources further downstream in new_sources and forced_extraction

These changes mean that the function now returns a persisted dask dataframe instead of a pandas dataframe, so I removed the subsequent lines in main.py which convert this dataframe to dask.

[ddobie]

The code looks good. I've just tried to put a run through and I'm hitting an error, but it doesn't seem to be related to the changes - will try and take a look at it in detail soon.