Add alternate implementation for split_dataframe_by_prefix()

[ptgolden]

Replaces #609. I started from scratch with a total rewrite rather than adding to the commit history of that one. The output here is identical, but it's hopefully easier to follow.

The point is the same: when splitting a MappingSetDataFrame by l subject prefixes, m object prefixes, and n relations, iterate over msdf.df once rather than l * m * n times.

[cthoyt]

here's a more straighforward way for iterating over the subset of columns you want:

Suggested change

	for mapping in msdf.df.itertuples(index=False):
	group = SSSOMSplitGroup(
	parse_curie(getattr(mapping, SUBJECT_ID), strict=True).prefix,
	parse_curie(getattr(mapping, OBJECT_ID), strict=True).prefix,
	parse_curie(getattr(mapping, PREDICATE_ID), strict=True),
	)
	for subject_id, predicate_id, object_id in df[[SUBJECT_ID, PREDICATE_ID, OBJECT_ID]].values:
	group = SSSOMSplitGroup(
	parse_curie(subject_id, strict=True).prefix,
	parse_curie(object_id, strict=True).prefix,
	parse_curie(predicate_id, strict=True),
	)

[ptgolden]

Will change. I used itertuples() on purpose because I thought it would save some memory/processing, but after profiling- it doesn't!

[ptgolden]

Ah, I remember the other reason I did it. In the next line, I'm storing the whole row (as a NamedTuple) in a list (mappings_by_group[group].append(mapping)). That list is parsed into a dataframe later:

    for group in expected_split_groups:
        ...
        mappings = mappings_by_group.get(group, None)
        ...
        split_to_msdf[split] = from_sssom_dataframe(
            pd.DataFrame(mappings), prefix_map=dict(subconverter.bimap), meta=meta
        )

[cthoyt]

@ptgolden I just added a test case to make sure we don't break backwards compatibility with handling of CURIEs that aren't registered properly in the internal converter. We'll also want to address this

[cthoyt]

I just merged all of the code together so we could get a better look at the time:

$ wget https://github.com/monarch-initiative/mondo/raw/refs/heads/master/src/ontology/mappings/mondo.sssom.tsv
$ mkdir results

$ time sssom split mondo.sssom.tsv -d results
________________________________________________________
Executed in   14.16 secs    fish           external
   usr time   14.16 secs    0.19 millis   14.16 secs
   sys time    3.49 secs    2.25 millis    3.48 secs

$ time sssom split mondo.sssom.tsv -d results --method patrick
________________________________________________________
Executed in   12.76 secs    fish           external
   usr time   12.80 secs    0.20 millis   12.80 secs
   sys time    3.44 secs    2.35 millis    3.43 secs

$ time sssom split mondo.sssom.tsv -d results --method disjoint-indexes
________________________________________________________
Executed in   12.74 secs    fish           external
   usr time   12.68 secs    0.12 millis   12.68 secs
   sys time    3.59 secs    1.49 millis    3.58 secs

$ time sssom split mondo.sssom.tsv -d results --method dense-indexes
________________________________________________________
Executed in   12.62 secs    fish           external
   usr time   12.64 secs    0.19 millis   12.64 secs
   sys time    3.51 secs    2.29 millis    3.50 secs

It might also be worth doing a memory profiling. If they're all comparable, then I would definitely opt for more idiomatic, reusable code rather than using a new custom implementation. @ptgolden can you run the same profiles I did above and let me know what it looks like?