Update to icu4j 78.2 and Unicode 17

[mpilquist]

Identified a few issues in the existing tests that this PR fixes:

• CodePointMapper can return strings that are not canonicalized whereas icu4j, as part of further IDNA processing, returns canonicalized strings -- e.g., "궈ㄻ" is returned as code points [44424, 4529] whereas icu4j returns 44434. This was fixed by adding an additional NFC normalization in the test suite.
• Differing normalization behaviors between java.text.Normalizer and icu4j when dealing with undefined unicode characters -- e.g., given the string "\u0360\u1ac6", Normalizer does not treat the undefined character \u1ac6 as a diacritic despite being in a block reserved for diacritics, whereas Normalizer2 does. This was fixed by adding a check for undefined characters when asserting equivalence.
• Differences in handling modifiers without base characters when normalizing input -- e.g. given the string "\u0345\u20e5", NFC normalization reorders the code points to "\u20e5\u0345" which then results in a different input to the mapping step than what icu4j is given. This was added as an "inconsistency" check. This represents a whole class of such errors that can occur randomly with the Scalacheck input, but occur infrequently enough to not be a major nuisance.

This PR also addresses a few issues with the build structure:

• Incremental compilation speed is improved by caching generated sources. Changing the unicode version or running a clean build will result in re-generation. This significantly improves compilation time and further allows incremental compilation without a network connection (b/c every codegen fetches mapping files from unicode.org).
• Fixes unused import warnings when launching SBT.