Currently watching files uses polling which is inefficient with large
codebases. On my laptop `./mill --watch` uses 20% CPU just to watch the
files for changes.

This uses `oslib`'s `watch` module to watch for the file changes instead
of polling them.

We had to update oslib (com-lihaoyi/os-lib#386)
to:
- add a capability of filtering what folders it watches. This is needed
because watching large generated folders like `out` or `.bloop` emits
Java FS watcher `OVERFLOW` events.
- remove the random `println`'s that `oslib.watch` had.

The current approach to watching is to watch the workspace root via FS
watching. FS watching only works with folders and we have to watch
`root/build.mill`, and thus, `root/`, so everything else falls under it.

Mac OS watcher performs watching recursively natively, but on linux
oslib adds recursive watches itself, so we use `oslib.watch` filter to
prevent watching anything that is not an ancestor of watched root. For
example, if we have source at `root/module-a/src/`, we'll watch `root/`,
`root/module-a/` and `root/`.

Finally, events are filtered so that unrelated changes (like creating
`root/random-file.txt`) does not trigger reevaluation.

The fs watching is enabled by default and can be disabled via
`--watch-via-fs-notify=false`.

`0.12.x` version: #5073

---------

Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>