fix(convert): move a mark's edge whitespace outside its delimiters

[willkg]

Fixes #204.

read/export trimmed the whitespace at the edge of a bold, italic or strikethrough span and never put it back. So <strong>bold </strong>next read as **bold**next and published as one word, as on the incident table in #204 ("features for13.5h"). The editor leaves a space typed at the end of a bold run inside the run, so this is a common shape.

What changes

• Marks (renderMark): whitespace at a mark's edge, Unicode whitespace included, moves outside the delimiters: **bold** next. It can't stay inside, because CommonMark refuses **bold ** as emphasis.
• Wrappers with no Markdown syntax (a coloured <span>, <u>, <sup>) no longer trim their edges. Without this, coloured bold text would still read as **bold**next.
• Link text keeps its edge whitespace inside the brackets ([see ](url)here). Markdown allows that and publishes it back.
• Neighbours that came up while verifying:
  • A mark holding only whitespace rendered as ****; it now renders as the whitespace.
  • A <br /> at a mark's edge was dropped; it now moves out with the rest.
  • a <br /> read as three trailing spaces and then two, so the Markdown wasn't a fixed point.
  • A hard break in a heading ended the heading; it stays a literal <br />.
• Where moved whitespace meets whitespace beside the mark, one space is kept, since storage would collapse the second on the next publish.

Testing

• New storage2md/mark-edge-whitespace case: the issue's three rows, a leading space, a no-break space, nested marks, a whitespace-only mark, breaks at a mark's edge, a coloured <span> and <u>, links, headings, and the incident table cell. It also runs through the round-trip fixed-point check.
• The split-marks golden and TestStorageToMarkdownCoalescesSplitMarks had recorded the bug (**a**[b… for <strong>a </strong><a>b…); both are corrected.
• Table property test:
  • The generator now puts a space at a mark's edge and follows such a mark directly with text. The edge spaces come from a separate random stream, so every existing seed's table is otherwise unchanged.
  • The model treats a space at a mark's edge as the same inside or outside, and two runs of one mark separated only by a space as one run.
  • Against main's converter it fails at seed 5; with this PR, all 3000 seeds pass, as did a 60s fuzz run (~200k execs).
• make check passes.

Left for later

• read: two hard breaks in a row become a paragraph break #213: two hard breaks in a row read back as a paragraph break. This already happens on main. Since this change a <br /> at a bold run's end followed by another <br /> hits it too, where before the first break was simply dropped.
• read: an empty paragraph before a list in a table cell is lost #214: an empty paragraph before a list in a table cell is lost. Found by the property test when the generator's seeds moved; unrelated to marks.
• A mark whose text starts or ends with punctuation next to a letter outside it (a**(b)**) still fails CommonMark's other flanking rule and reads back as literal asterisks. This PR doesn't change that behaviour.