More large-catalog speedups

[jvansanten]

This PR introduces a bundle of changes, that, together with #396, reduce the time to initialize an injector and run a single trial with 10k sources and the 10-year NT dataset (gamma_precision reduced to 0.25, so only 8 gamma points) from 6.2 hours to 3 minutes. Of that, a solid 2 minutes are spent in load_data()*. The speedups arise as follows:

• Add a new TableInjector that keeps its MC data in an astropy.table.Table, sorted by declination, rather than in a gigantic structured array. This helps in two ways. First, operations on individual columns of the table make better use of caches, since the values are stored contiguously in memory, rather than 176 bytes apart as in the array. Second, sorting the table by declination means that the expensive band-masking step can be replaced by a cheap binary search. This is also more cache-efficient than indexing with a boolean array, since applying a boolean mask requires reading every single value in the column (or worse, the structured array). This injector is in all ways better than LowMemoryInjector (and perhaps, the base MCInjector)
• Optimize StdMatrixKDEEnabledLLH's coincident event selection in much the same way, using a Table to store the input data, sorted in declination, and processing sources in declination order.
• Build up SoB_only_matrix row-by-row in canonical CSR format, and skip constructing an explicit coincidence matrix entirely. Updating a scipy.sparse matrix turns out to be surprisingly expensive, and largely unnecessary.

*Most of the remaining fat is in flarestack.icecube_utils.dataset_loader.load_dataset() and its use of append_records(). This could be obviated by using astropy Table everywhere, but would make the most sense if the data were stored that way in the first place. That's probably too large of a change at this point.

[sathanas31]

Finally tested! For same (single) scale & seed, 1 trial for 1 src, the scrambled dataset using table_injector is the same with that from low_memory_injector. Using this table_injector dataset, the llh kwargs were compared and found same with what you get from main (merged #396). Unless you want me to check low-level params (eg coincident data), we good to go :D

[JannisNe]

If the datasets are the same, then also the coincident data will be.