chore(core): simplify input assumptions for actions_normalize

Per discussion in #10422, we can assume that input cached_context is
always NFD. However input actions->output may not start at a
normalization boundary, so we still need to backtrack to a normalization
boundary in order to get our NFC output. But cached_context never need
change.

This makes no change to the algorithm, but tweaks some of the unit tests
to adhere to this input assumption.

Note: we could consider adding a debug assertion that cached_context is
NFD.