Use a set for TargetPython.get_tags for performance

[hauntsaninja]

This makes a representative pip-compile about 1.6x faster

# Before
λ hyperfine -w 1 -M 3 -p 'rm req.txt || true' 'python -m piptools compile req.in --output-file req.txt --strip-extras --resolver=backtracking'
Benchmark 1: python -m piptools compile req.in --output-file req.txt --strip-extras --resolver=backtracking
  Time (mean ± σ):     28.552 s ±  2.552 s    [User: 20.477 s, System: 1.547 s]
  Range (min … max):   26.446 s … 31.390 s    3 runs

# After
λ hyperfine -w 1 -M 3 -p 'rm req.txt || true' 'python -m piptools compile req.in --output-file req.txt --strip-extras --resolver=backtracking'
Benchmark 1: python -m piptools compile req.in --output-file req.txt --strip-extras --resolver=backtracking
  Time (mean ± σ):     17.570 s ±  0.450 s    [User: 11.537 s, System: 1.345 s]
  Range (min … max):   17.095 s … 17.989 s    3 runs

This makes the set.isdisjoint operation done over here much cheaper: https://github.com/pypa/pip/blob/593b85f4a/src/pip/_internal/models/wheel.py#L92

The reason for this is because a Python usually has a lot of supported tags. The Python I used above has 2000 supported tags! Whereas a wheel usually only has one or two file tags.

The CPython code will unfortunately iterate over all 2000 tags to check if there's a match. Only if the other collection is a set will CPython think to swap the operands to iterate over the shorter collection: https://github.com/python/cpython/blob/35963da40f/Objects/setobject.c#L1352

[hauntsaninja]

Oh, also note I tried narrowing the type on Wheel.supported to only take Set[Tag] instead of Iterable[Tag]. It's pretty doable, let me know if I should pursue this (could also be a follow-up PR).

[uranusjr]

I don’t see why we need to narrow it though. Iterable seems just fine to me.