Fix font matrix handling for an edge case

[LaurenzV]

By pure chance, I found that Typst main fails to properly render some text with the new subsetter:

#set text(font: "PingFang SC", fallback: false)

你

After debuging, I realized that this is because, if the TOP DICT matrix is empty and the FONT DICT matrix is not, then in the subsetted font we would write 0.001 for the TOP DICT matrix and the existing FONT DICT matrix, which means that we are applying two transforms now. After rereading the comment in the ghostscript tracker, what ghostscript does is initialize TOP DICT matrix to 0.001 like we were doing before, but scales the FONT DICT matrix by x1000.

The way it is in the PR is not entirely correct because I don't think you can just scale the matrix like that, but I haven't seen any font matrix that has a skew/translation part, so I hope this should be fine for now.

Before (current Typst main): Only Apple Preview would render the PDF:

After: All viewers render it, including Acrobat:

Having no TOP DICT matrix but a FONT DICT matrix seems to be somewhat of an edge case, which is also why no tests are affected by this. So hopefully this doesn't cause any unwanted regressions.

[LaurenzV]

Alright, so this PR is now based on #3, so it shouldn't be merged before that.

I've created a "synthetic" font now with two crazy matrices and compared the output to what ghostscript would produce with such an embedded font, and they match (apart from the fact that our 0.799995 turns to 0.8 for them). And I also manually verified that the original font works in Typst-created PDFs. So hopefully this implementation is better now.