chore: use if TYPE_CHECKING: blocks

[mayeut]

This will become the default in Python 3.14 in any case given the current lint rules.
It ensures typing only imports do not slow down execution (admittedly only useful in edge use cases, such as querying help).

If #2797 makes it, it will further speed-up this one:

hyperfine --warmup 10 \
     -n "main"                --prepare "git checkout main"               ".nox/tests-3-15/bin/python -m cibuildwheel --help" \
     -n "Henry's PR"          --prepare "git checkout pr/2797"            ".nox/tests-3-15/bin/python -m cibuildwheel --help" \
     -n "type-checking"       --prepare "git checkout type-checking"      ".nox/tests-3-15/bin/python -m cibuildwheel --help" \
     -n "type-checking-lazy"  --prepare "git checkout type-checking-lazy" ".nox/tests-3-15/bin/python -m cibuildwheel --help"
Benchmark 1: main
  Time (mean ± σ):     308.7 ms ±  16.7 ms    [User: 194.7 ms, System: 44.1 ms]
  Range (min … max):   294.3 ms … 349.3 ms    10 runs
 
Benchmark 2: Henry's PR
  Time (mean ± σ):     106.2 ms ±   9.4 ms    [User: 62.7 ms, System: 17.6 ms]
  Range (min … max):    88.7 ms … 118.2 ms    19 runs
 
Benchmark 3: type-checking (this PR)
  Time (mean ± σ):     306.9 ms ±  15.5 ms    [User: 194.2 ms, System: 43.2 ms]
  Range (min … max):   290.2 ms … 342.8 ms    10 runs
 
Benchmark 4: type-checking-lazy
  Time (mean ± σ):      77.5 ms ±   6.1 ms    [User: 45.6 ms, System: 13.7 ms]
  Range (min … max):    68.9 ms …  86.3 ms    22 runs
 
Summary
  type-checking-lazy ran
    1.37 ± 0.16 times faster than Henry's PR
    3.96 ± 0.37 times faster than type-checking (this PR)
    3.98 ± 0.38 times faster than main

[henryiii]

Sounds good to me.