Optimisations and bugs in SpellChecker

A bug was introduced in a recent refactoring: correct words didn't have their
SpellCheckSpan removed, leaving a lot of useless spans.

SPAN_EXCLUSIVE_EXCLUSIVE should never have a 0-length. With Japanese characters
wordStart could be equal to wordEnd when parsing the text: skip these.

Using toString().substring(...) instead of subSequence(...).toString() which
is more efficient.

Change-Id: I670870a34565939b676400091f4852152a7f7124