[salkinium]

Follow-up to #22: instead of emulating the loose character box of the old pdfium (float32 rounding and all, meh), this computes our own box from what the layout heuristics actually want: the font size and the advance width, placed by the full text matrix.

• Character._em_box() is one font size tall and one advance width wide, which also works for rotated text, so Page._fix_bboxes() and its cache are gone.
• Lines of the same font size are merged into the longer line instead of comparing their heights for equality, which depended on rounding errors.
• Lines of the same font size whose characters are printed on top of each other stay separate lines (numerator and denominator of a formula, overlapping index entries). This resolves the FIXME about formulas colliding vertically.
• Regression test pages are byte-identical.
• Converted the 220 documents whose archived HTML changed in the last year: 2441 of 2833 chapter files are identical to [pdf] Upgrade to pypdfium2 v5 #22, 1219 hunks differ in 142 documents.

What changes in the output:

	#22	this PR
Headings starting with <sub>	468	208
Interleaved sub/sup garbage runs	32	33

The remaining 208 are real subscripts like V<sub>BAT</sub>. Fractions now come out as numerator / dashes / denominator on separate lines instead of a <sub>/<sup> salad.

Missing:

• Full regeneration of the HTML archive, hence the draft. Only those 220 documents were converted so far.
• Fractions are readable but not structured, and a denominator sometimes ends up at the end of the paragraph instead of under its bar.
• A wrapped mm² in a table cell is still wrong, just differently.
• General gap-based space insertion: fixes Armdouble but breaks Cortex®-M7 into Cortex -M7 since the ® is filtered out, so only the bullet points get their space back for now.