Style and document cleanup — remove unused styles, numbering and media #115

Once #114 can say which styles are reachable, the useful operation is dropping the rest.

BaseStyle.delete() already exists, so this is not about removing one style — it is about removing the right set safely, and about the neighbouring dead weight that travels with it.

The motivating number: a document created by this library carries 164 style definitions and 137 latent-style exceptions. A one-paragraph document references one style. word/styles.xml is 349 KB of the 826 KB uncompressed total.

Shape

document.styles.remove_unused()                       # prune to the reachable closure
document.styles.remove_unused(keep=["Quote", "Caption"])   # plus ones you plan to apply
document.styles.latent_styles.trim()                  # drop the lsdException list

• Styles.remove_unused(*, keep=(), keep_defaults=True) — deletes every style outside the reachability closure from Style usage analysis — which styles a document actually uses #114, plus anything named in keep and its own closure
• Returns what it removed, so the caller can report it

Beyond styles — the same cleanup family

Worth deciding whether these ride along or get their own issue. They are the same "this document carries things nothing points at" problem:

• Unused numbering definitions. word/numbering.xml is 5.5 KB in a document with no lists. An w:abstractNum no w:num points at, and a w:num no w:numPr points at, are both dead. Make numbering definitions writable - define a list from scratch #98 is adding the write side of numbering, so the machinery to walk it is arriving anyway.
• Orphan media. An image part related from nothing, left behind when a run or a floating shape was deleted. The .delete() API added in 2.0.0 makes this reachable from ordinary use — deleting a paragraph containing a picture leaves the image part in the package.
• Unused fonts. word/fontTable.xml lists fonts no w:rFonts names. Cosmetic; low value; mentioned only so it is not rediscovered later.

A single Document.cleanup() that runs the lot, with flags, is probably the shape people want. Keep the individual operations public too — "remove unused styles but leave my media alone" is a reasonable thing to want.

Watch out for

• This is destructive and the closure is the only thing standing between it and a broken document. If Style usage analysis — which styles a document actually uses #114 misses an edge — a style referenced only from a header, only as a w:basedOn, only from w:lvl/w:pStyle in numbering — remove_unused() silently produces a document whose formatting has changed. Land the analysis with its tests first, and make the closure the single implementation both features use.
• Never remove the w:default="1" styles for any type, whatever the closure says.
• Never remove Normal. Word repairs a document that lacks it, and the repair dialogue is worse than the bloat.
• Deleting a style that other styles are basedOn must either be refused or must repoint the dependants. Silently orphaning them changes their formatting, because basedOn pointing nowhere means "based on docDefaults".
• Latent styles are a behaviour declaration for styles that are not defined — removing an lsdException changes whether the style shows in Word's gallery, not the document's appearance. Different risk profile from removing a definition; keep the two operations separate.
• Byte-reproducible output (tests/test_reproducible.py) must still hold after a cleanup pass.

Note on scope

This does not do anything to the bundled template — that is #113. This is an operation callers run on their own documents. The two overlap only in that the bundled template is the most obvious thing to point it at.