This adds an option to parseXmlDocument and parseXmlFragment to treat
CDATA sections in the input as normal text, outputting (normalized) Text
nodes instead of CDataSection nodes.

This re-implements the algorithms used to determine what prefixes to
assign and what declarations to output for them during serialization.
The algorithms in the DOM parsing and serialization spec contain too
many bugs that have so far not been addressed, and contain complicated
branching that makes it hard to reason about them.

In the replacement I've tried to remain close to the original behavior
in spirit. That is, existing prefixes (or, for elements, an inherited
default namespace) are preferred over authored prefixes but only if they
are still in scope. Authored prefixes are preferred over generated ones.
A new prefix is generated only if the authored prefix of an attribute
conflicts with a declaration on the same element. Declaration attributes
in the DOM are only preserved if they actually represent a change to the
namespaces in scope for their element and do not cause conflicts - the
spec preserved default namespace declarations in a few other cases but
those seemed unnecessary.

This fixes attributes being allowed to adopt the default namespace (with
empty prefix) and ancestor declarations not being considered as
candidates.

Looks like I missed this in the earlier implementation. This refactors the implementation a bit into a shape that also applies to any entities expanded as part of normalizing attribute values, both normal and default.

Jest isn't compatible with version 3 yet...

Fixes #151