In HTML documents that have been served with an XML mime type,
`document.evaluate` handles element names differently. In HTML documents
the XPath segments do not require prefixes on element names. In XHTML
documents however they do. Since the client always generates the same
un-prefixed XPaths regardless of document type, evaluation always
failed. There was a fallback path but it was only executed if
`document.evaluate` threw an error, not if it returned `null` as in this
case.

This commit resolves the issue by first attempting to evaluate the XPath
using custom logic that only handles simple XPaths and then using
`document.evaluate` only for more complex XPaths. The _simple XPath_
logic behaves the same in HTML and XML documents and ignores namespaces.
As a result it works with the XPaths that the client generates
regardless of document type and also regardless of whether the XPath
references only HTML elements or elements from other namespaces (eg.
MathML or SVG).

We could change the way that the client generates XPaths in future in
XML documents or SVG/MathML content within HTML documents, but the
client would still need to handle `RangeSelector` selectors on existing
annotations.

For more details, see https://github.com/hypothesis/client/pull/2590#issuecomment-702132563

Fixes #2592