feat(form_fit): set tabs on a paragraph's own tab stops

[airmang]

Stacked on #470 (default tab stops); the diff against that branch is this change alone.

What

FormFit and the page estimate now lay a tab out on the stops of the paragraph's tab definition, as Hancom Office does:

• The first stop past the tab whose text can be aligned there comes first. A left stop starts the text there, a right stop ends it there, a centre stop centres it there, and a decimal stop puts its last full stop there.
• The aligned text neither starts before the tab nor passes the line's end. A stop it cannot take is passed over. A left stop is taken where it stands, but past a stop already passed over, it too ends its text at the line's end.
• The automatic tab at the end is a right stop at the line's end, among the others. The automatic tab of a hanging indent is a left stop where its lines start.
• Then the default stops. Past a stop it could not take, a tab whose text has no room on the line after the default stop takes no width.
• A tab reaching the margin itself, even one of no width, goes on to the next line.

Before, such a tab was laid out as a space-class glyph.

Tests

• New Hancom-saved fixtures (saved by Hancom Office 2024, scrubbed) of paragraphs with stops of their own after labels, numbers and words. They cover a centre stop then the default stops, a left stop past the line's end, a decimal stop passed over, the automatic tab at the end and a right stop with no room. Without the line caches they fail before this change and pass after.
• test_a_tab_reaches_the_paragraphs_own_stops_first covers each rule of tab_reach.

🤖 Generated with Claude Code

[airmang]

Reviewed the custom tab-stop implementation and surrounding line measurement code, including the new 7c2e2e8 changes for left stops after a skipped stop and a zero-width tab at the margin. After parent #470 merged, retargeted this stacked PR to main and verified the resulting scope. On this exact new head integrated with current main dae1ee2, the tab-related FormFit/page-estimation suite passed: 792 passed, 733 deselected. No reproducible defect found in the reviewed scope. Holding merge because the new head's Python 3.10–3.14 CI jobs are still running; the successful checks and broader 1521-test run on the superseded c02e0d0 head do not certify the new head. Recheck head, base, threads and all applicable checks before merging.