perf(import): load tool submodules on demand and defer ingest metadata

[airmang]

First HwpxDocument.to_bytes() loaded every tool module through hwpx.tools.__init__, including exporters, viewers, diff, and mail merge. Tool exports and the previously reachable submodules now resolve on first access through static imports; save loads only the package validator and its required document validator. The Markdown converter also imports importlib.metadata only when it looks up the distribution version.

Existing exports, __all__, dir(), star imports, object identities, and static types are preserved. A targeted module setter preserves the doc_diff function when Python attaches its same-named child, including direct-child-first imports. Save validation remains intact.

Fixes #386.

Validation

• Full suite: 5769 passed, 19 skipped, 1 xfailed; coverage gate passed. Final import/diff rerun: 22 passed, including 12 new cold-process cases. Five import-cost checks failed before the change.
• Ruff, mypy (120-file scope plus both changed modules), pyright, typing scope, product boundary, public hygiene, error census, capabilities verification, and size ratchet pass.
• Rebuilt wheel installed in an isolated environment with only lxml: deferred imports, all exports, doc_diff import ordering/function replacement, version lookup, and save/reopen pass with warnings treated as errors.
• Three externally authored government documents (tests/fixtures/m3_gongmun_gold/NOTICE.md): all six unchanged/edited save-reopen cases pass, with every saved ZIP member byte-identical to main for the same operation.

Local measurements

11 interleaved fresh isolated processes per tree, identical Python 3.13 interpreter/dependencies, bytecode warmed once; medians:

	main	change
import hwpx	59.0 ms	58.0 ms
First to_bytes()	21.9 ms	15.8 ms
Tool modules after first save (including the package)	12	3
importlib.metadata after plain import	loaded	deferred

The hwpx-module import bound remains 97. Timings are observations on this machine.