We applied the same treatment for the sanity tests, where we ignore the
first line (the tailwind header).

An odd issue I have found is that diffing of big css files is _very_
slow. When te tests pass, then the first test takes `3302 ms`, however,
when it fails it takes `477482 ms` on my machine. That's almost 8
minutes.

* Add failing test for #4823

* Fix important variant `@apply`

Co-authored-by: Brad Cornes <[email protected]>

* update (old) incorrect snapshots

The version changed, so it is save to update these.

* add `shadow` so that we can test `@tailwind base`

We were only using `font-bold`, but that doesn't require any base rules.
We try to only inject rules that are required, even for `base` related
rules.
Adding `shadow` will ensure that we can test that `base` rules have been
added.

The trackedModified call in the tracking logic has a bug that
incorrectly removes `null` from pathnames

Currently:
```
let pathname = parsed.href.replace(parsed.hash, '').replace(parsed.search, '')
```
Where if `parsed.hash` or `parsed.search` are missing (which is
mostly the case for FS files), the value is null
eg `{ hash: null, search: null}`
  - In which case, we essentially trim `null`
converting `/mypath/nulldir/file.js` -> `/mypath/dir/filejs` and
breaking builds (see #4920 ).

Fix checks if `hash` or `search` are set before replacing them

Fixes #4920

* pin clean-css to version `5.1.4`

* update fixtures