GH-104996: Implement path joining algorithm in pathlib

[barneygale]

Copy the ntpath.join() algorithm into pathlib and adjust it to remove string concatenation. The resulting drive, root and tail are stored on the path object without creating an intermediate joined path.

Timings in microseconds:

	Before	After	Change
PurePosixPath().root	1.13	1.03	n/c
PurePosixPath("/a").root	1.93	1.99	n/c
PurePosixPath("/a", "b").root	2.81	2.45	15% faster
PurePosixPath("/a", "b", "c").root	3.29	2.86	15% faster
PureWindowsPath().root	1.1	1.02	n/c
PureWindowsPath("/a").root	2.11	2.14	n/c
PureWindowsPath("/a", "b").root	3.62	2.79	30% faster
PureWindowsPath("/a", "b", "c").root	4.26	3.31	30% faster

• Issue: Optimize joining of pathlib.PurePath() arguments. #104996

[barneygale]

@eryksun do you think you could you review this please? It's a variant of the ntpath.join() algorithm, but we skip joining parts in the tail, and skip combining the the drive, root and tail.

[eryksun]

I'm not thrilled about duplicating the code from ntpath.join(). Have you considered adding an ntpath.splitseq(sequence) function that returns (drive, root, path_list)? Then ntpath.join() could be implemented as follows:

defjoin(path, *paths):
paths= [os.fspath(path), *paths]
ifisinstance(paths[0], bytes):
path=b''sep=b'\\'seps=b'\\/'colon=b':'else:
path=''sep='\\'seps='\\/'colon=':'try:
drive, root, path_list=splitseq(paths)
forpinpath_list:
ifpathandpath[-1] notinseps:
path+=seppath+=p# If needed, add a separator between a UNC drive and path.ifpathandnotrootanddriveanddrive[-1:] !=colon:
returndrive+sep+pathreturndrive+root+pathexcept (TypeError, AttributeError, BytesWarning):
genericpath._check_arg_types('join', *paths)
raise

[barneygale]

The only problem I see there is that os.path.splitseq() might be quite difficult to explain in documentation!

[eryksun]

The only problem I see there is that os.path.splitseq() might be quite difficult to explain in documentation!

I think a better function name would help. Maybe mergepaths() is better.

path_list should conserve the slashes in each segment. Empty pathname strings should be retained.

[barneygale]

Maybe partialjoin()? After all, it's 80% of the join() method, and probably easiest to explain as "like join(), but returns a 3-tuple"

[barneygale]

I'm withdrawing this PR as it bakes elements of pathlib's current normalisation logic into the path parsing/joining, and the overlap precludes user customisation and some other optimisations I have in mind.