Keep score_cutoff of normalized scorers in double precision

[raffaelemancuso]

Fixes #494.

get_score_cutoff_f64 in src/rapidfuzz/cpp_common.pxd stored the cutoff in a Cython float, which is a single-precision C float, although it returns a double and every caller passes a double on to rapidfuzz-cpp. A cutoff whose float32 value rounds up (0.8 becomes 0.800000011920929, likewise 0.925 and 4/7) then rejected a score exactly equal to it:

from rapidfuzz.distance import Indel
Indel.normalized_similarity(list("abcde"), list("abcdZ"), score_cutoff=0.8)
# 0.0 before, 0.8 after (and 0.8 in the pure Python implementation)

The first commit declares c_score_cutoff and the worst_score / optimal_score parameters as double.

The regression test, test_score_cutoff_equal_to_score in tests/distance/test_distance.py, checks that every normalized scorer returns a score equal to score_cutoff rather than 0 or 1, for both normalized_similarity and normalized_distance, with the C++ and the pure Python implementations alike (through the existing GenericScorer). It also caught a second, smaller divergence, fixed in the second commit: the pure Python Prefix and Postfix computed normalized_distance as 1.0 - normalized_similarity, which can differ in the last bit from distance / maximum, the value the C++ implementation returns (0.19999999999999996 instead of 0.2); with that value as score_cutoff the two implementations disagreed. They now compute it as the C++ implementation does (and as Levenshtein_py already did).

Tests

• Against the released 3.14.6 wheel, the new test fails for 9 of the 10 scorers (Postfix scores 0 on these strings).
• Built from this branch on Windows 11 x64 (MSVC 19.51, CPython 3.14.5): 412 passed, 19 skipped (numpy not installed).
• CI on my fork: the test suite passes (431 passed) in "Coverage of Test Build", whose final Codecov upload fails there for lack of a token; "Submodule Test" fails as it does on main (submodule v3.9.0 against v4.1.0).

I also added two entries under an [Unreleased] heading in CHANGELOG.rst; happy to drop them if you prefer to write the changelog at release time.