fix: tighten German normalization edge cases

[apples-kksk]

This is a small follow-up patch for the German G2P PR at hexgrad#97.

It tightens a few normalization edge cases:

• avoid consuming Uhr inside longer words like Uhrzeit
• leave invalid times such as 25:00 Uhr and 23:99 Uhr unchanged instead of expanding each number around the colon
• use Decimal for currency amounts so values like €9,999 round cleanly to zehn Euro rather than neun Euro und einhundert Cent

Verification:

• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. uv run --isolated --no-cache --no-project --python 3.11 --with pytest python -m pytest -q tests -p no:cacheprovider -> 67 passed, 4 skipped
• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=. uv run --isolated --no-cache --no-project --python 3.11 --with pytest --with phonemizer-fork --with espeakng-loader python -m pytest -q tests/test_de.py -p no:cacheprovider -> 71 passed
• python -m compileall misaki/de.py tests/test_de.py -> passed
• git diff --check -> passed

I also noticed pyproject.toml adds the de extra, but uv.lock does not yet include it. I could not regenerate the lockfile cleanly because the existing he extra fails resolution for Python 3.12/3.13 via mishkal-hebrew>=0.3.2, so I left the lockfile untouched here.

[semidark]

Overall this change set looks correct and test coverage for the edge cases is strong. I left two non-blocking suggestions focused on simplifying invalid-time handling and avoiding potential placeholder-token collisions.