Add `linter_level` to `Linter()` metadata.

[AshesITR]

          Here's what cost > 1s in `lint()` during `lint_package()` currently:

We can see 30% of the time is spent in xml_find_all().
What surprised me was is_lint_level() taking a whole 8s. That may be worth converting to an attribute of Linter()?

Originally posted by @AshesITR in #2350 (comment)

[AshesITR]

~10% of the time lint_package() took (80s) is spent in is_lint_level().
We could avoid this by adding an attribute to Linter() that specifies lint_level.

For maximum backward compatibility, I suggest it to default to NA_character_ keeping the current behavior.
If it is set to "file" or "expression", only those source_expressions should be passed into the linter.
This means we can drop the check for all our linters and delegate it to lint() checking linter attributes.

[MichaelChirico]

I've actually had in mind something along these lines, namely, that most linters could be toggled very easily between file- and expression-level operation, with the difference maybe being a slight difference in the XPath (but in most cases the same XPath works for both cases).

It would be nice to have an easy API through lint() to say "I don't care about expression-level cacheing, just run file-level lints".

[AshesITR]

Working on this atm (diff will be large because it needs to touch all linters for the performance benefits)