Fixes a bug where we were too eager to remove empty containers.
We already had logic to skip removing empty containers if it had the
filter attribute, which is needed to apply a filter to the whole area.
However, the filter can also be defined through CSS. We did not
properly handle this case, and treated the node as if it had no filter
at all.
This computes the styles and checks the stylesheet as well. (We also
move the logic down to avoid computing the styles eagerly.)
When running this plugin on an SVG with an empty block in the CSS, it
would apply it to the matched elements by adding an empty `style`
attribute. See the test for an example.
This resolves that by just dropping the declaration if it's empty.
Style property names are case-insensitive, for example `fill`, `FILL`,
and `FiLl` are all valid but refer to the same property. So when
inlining styles, we compare the lowercase strings and only use the last
instance of the property rather than embed them all.
A lot of new sources of regression tests may come and it's important to
make debug simpler.
Now regression-extract.js downloads and write svg files into
test/regression-fixtures. regression.js run each svg in this folder.
Mismatched svg diff is written into test/regression-diffs.
Visual regressions tests will give us more confidence in SVGO stability.
Test cases are downloaded from w3.org, screenshooted with playwright and
compared with pixelmatch.
