Reasons to do this:

- setuptools is no longer the only game in town. When cibuildwheel was created, setuptools was the only way to make wheels. these days, there are many build backends so special casing setuptools doesn't make so much sense
- we don't need to preinstall `wheel` any more, setuptools will do that itself when it needs it.
- projects with pyproject.toml (i.e. the vast majority of projects going forward) don't benefit from the pin anyway, because both `pip` and `build` create an isolated environment to install `build-system.requires` into.
- this makes it easier for end users (such as @webknjaz) to control build-system versions using `PIP_CONSTRAINT`, because there's less potential for a user-specified version to conflict with one of our pins.
  - (long-term, I'd love `build` to [support this directly](pypa/build#292), and then we could have a proper way to keep pyproject.toml dependencies loose and forward-compatible while retaining build determinism. But PIP_CONSTRAINT is the best we have for now. PIP_CONSTRAINT isn't ideal because it affects other things, like the test virtualenv).

This PR also adds a pin for pypa/build. It appears that was missing before, but we should pin it.

…erge