gh-149079: Optimize sorting in unicodedata.normalize()

[serhiy-storchaka]

Sort the Py_UCS4 buffer instead of PyUnicodeObject. This allows to avoid the use of PyUnicode_READ() and PyUnicode_WRITE().

• Issue: O(n²) insertion sort in unicodedata.normalize("NFC") canonical ordering #149079

[serhiy-storchaka]

./python -m timeit 's=("a"+"\u0300\u0327"*1000)*100; from unicodedata import normalize' -- 'normalize("NFC", s)'

Baseline: 100 loops, best of 5: 3.76 msec per loop
This PR: 100 loops, best of 5: 3.57 msec per loop

./python -m timeit 's=("a"+"\u0300\u0327"*9)*10000; from unicodedata import normalize' -- 'normalize("NFC", s)'

Baseline: 100 loops, best of 5: 3.99 msec per loop
This PR: 100 loops, best of 5: 3.84 msec per loop

[eendebakpt]

@serhiy-storchaka On your benchmark I can improve from 3.75 ms to 2.0 ms by using a more efficient search in find_nfc_index. See main...eendebakpt:gh-149079-find-nfc-index. The changes in the PR look good at first sight.

[serhiy-storchaka]

by using a more efficient search in find_nfc_index.

Looks interesting. But this is a different issue, not directly related to #149079. Can you open a new issue?