WordDelimiterFilter produces invalid offsets in single word case [LUCENE-7795]

This problem is not present in WordDelimiterGraphFilter, but it is present in WordDelimiterFilter's interaction with HTMLStripCharFilter.

Test code:

```Java
public class TestTokenizationIssue2 {
public static void main(String... args) throws IOException {
HTMLStripCharFilter charFilter = new HTMLStripCharFilter(getText());
WhitespaceTokenizer whitespaceTokenizer = new WhitespaceTokenizer();
whitespaceTokenizer.setReader(charFilter);
// WordDelimiterGraphFilter wdgf = new WordDelimiterGraphFilter(whitespaceTokenizer,
// WordDelimiterGraphFilter.GENERATE_WORD_PARTS, CharArraySet.EMPTY_SET);

WordDelimiterFilter wdgf = new WordDelimiterFilter(whitespaceTokenizer,
WordDelimiterFilter.GENERATE_WORD_PARTS, CharArraySet.EMPTY_SET);
wdgf.reset();

while (wdgf.incrementToken()) {
CharTermAttribute charTermAttribute = wdgf.getAttribute(CharTermAttribute.class);
OffsetAttribute offsetAttribute = wdgf.getAttribute(OffsetAttribute.class);

System.out.println(charTermAttribute.toString() + " - " + offsetAttribute.startOffset() + ',' + offsetAttribute.endOffset());
}
}

private static Reader getText() {
return new StringReader("“Risk");
}
}
```

The offsets produced by the WordDelimiterFilter are 1,10. With WordDelimiterGraphFilter the offsets produced are 0,10. It should be 0,10 as this is the original text: ```
“Risk
``` - and 1 is between the ampersand and hash.

Inside WordDelimiterFilter, I believe the conditional branch from "if (isSingleWord && startOffset <= savedEndOffset) " is invalid and it should always use the saved start and end offsets because it can't make the assertion that the iterator's current and end are reliable markers.

---
Migrated from [LUCENE-7795](https://issues.apache.org/jira/browse/LUCENE-7795) by Michael Braun (@michaelbraun)

Start with WordDelimiterFilter and its interaction with HTMLStripCharFilter, using the Java reproduction in the issue. Compare its offsets with WordDelimiterGraphFilter; done means the single-word case reports offsets 0,10 for the original ““Risk” text and has regression coverage.

Written by the indexing model from the issue text.