It's currently implemented by delegating to SelectSingleNodes and returning the first one.  While the list is lazily-populated, this still entails creating an `XPathNodeList`, creating a `List<XmlNode>`, storing the enumerated into the list, and then returning the element from the list, which is then thrown away.  With just a few lines, we can cut through all of that.

No need for these to be allocated classes.  They can just live on the stack where they're created.

This shows up on the hot path of parsing the xpath expression.  We can use a span to avoid bounds checking.

NextChar is used by lots of routines to advance to the next character.  We can streamline it to avoid the bounds check when indexing into the string.

The typical case is there isn't any whitespace, so inline that fast check.  This was showing up as a few percentage of a simple scenario.

buyaa-n left a comment