fix: delete trailing ASCII when accepting prediction

[ensan-hcl]

Summary

• keep the existing predictive match behavior that ignores a trailing ASCII letter
• also delete that trailing ASCII letter when accepting the prediction candidate
• add regression tests for the trailing-ASCII and non-trailing-ASCII cases

Context

With "Enable predictive typing in development" enabled, accepting a candidate after input like おはようございm produced おはようございmます because the trailing ASCII letter was ignored for matching but no longer deleted on accept after the recent deleteCount refactor.

Testing

• swift test currently fails before reaching these tests due to an upstream dependency build error in AzooKeyKanaKanjiConverter (DictionaryBuilder.swift missing import OrderedCollections under Swift 6.2 member import visibility checks)

[ensan-hcl]

GA failure looked consistent with the AzooKeyKanaKanjiConverter revision pinned in Core/Package.swift.

I updated it to e57b2c82cdecdeb8dc9058a8d04264e79bea1ab3, which should include the Swift 6.2 compatibility fix for the previous build break.

Verification on this branch:

• swift test in Core/ now completes successfully

This is included in commit 3b2d420.