When normalizing, we need to stop processing on an NFC boundary, not an
NFD boundary, to support normalizations such as in Bengali, where
appending `U+09D7` to a context of `U+0995 U+09C7` should result in
`U+0995 U+09CC`.
The specification is unclear on this; see https://unicode-org.atlassian.net/browse/CLDR-19218
This also updates the ldml keyboard unit test suite to support running
in full NFC mode (used in all Engine implementations) as well retaining
the NFD mode (now only used by the debugger).
Side note: the Bengali normalization failure case was picked up by the
improvements to the unit test suite, proving once again that good tests
are so valuable.
Fixes: #15491Fixes: #15505
Follows: #15488
Relates-to: CLDR-19218