gh-117865: Speedup import of inspect module

[danielhollas]

This is a revived version of #119526 (with Alex's blessing), using the new fancy lazy import machinery to defer imports of re and tokenize modules, for a ~20% speedup.

Before

❯ hyperfine -w 5 "./python -c 'import inspect'"
Benchmark 1: ./python -c 'import inspect'
  Time (mean ± σ):      19.2 ms ±   1.9 ms    [User: 14.4 ms, System: 4.5 ms]
  Range (min … max):    16.3 ms …  23.4 ms    152 runs

After

❯ hyperfine -w 5 "./python -c 'import inspect'"
Benchmark 1: ./python -c 'import inspect'
  Time (mean ± σ):      16.4 ms ±   2.0 ms    [User: 12.3 ms, System: 3.9 ms]
  Range (min … max):    13.8 ms …  22.2 ms    191 runs

Part of #137855

• Issue: Inspect module is prohibitively expensive to import #117865

[danielhollas]

Not sure about the convention here: where do we put lazy imports?

[JelleZijlstra]

My intuition would be that you put them in the same place you'd have put them before, just with "lazy" in front.

(Yay, a new thing to argue about! I guess isort will have to make the decision for us in the future.)

[danielhollas]

Thanks, yeah I put them back, it also made the diff nicer. (currently there doesn't seem to be consistent sorting in stdlib anyway)

[JelleZijlstra]

Makes sense. re is used only in one place in formatannotation, so many users won't need it. tokenize is used a bit more heavily but most of the API in the module still won't need it.