Fix double-encoded entity references flattened during XML round-trip

[ronaldtse]

When XML contains double-encoded entities (e.g. < to represent
literal text <), the round-trip through model serialization flattened
them by one level, producing semantically different or invalid XML.

Root cause: add_text_with_entities treated ALL entity-like patterns in
text content as EntityReference nodes, including standard XML entities
(lt, gt, amp, apos, quot) and numeric character references. This caused
&lt; to become < (i.e. <) after round-trip.

Three fixes in add_text_with_entities:

• Exclude standard XML entities from EntityReference creation; treat as
  text nodes so the serializer handles proper escaping
• Use #match instead of #match? for the entity name check (match? does
  not populate in Ruby, silently breaking the exclusion check)
• Tighten entity name regex to require letter-initial names per the XML
  spec, preventing invalid entity references like &1; from shell syntax

Non-standard entities (copy, nbsp, mdash, etc.) continue to be preserved
as EntityReference nodes as before.

Fixes: rfc8792, rfc8846, rfc9052, rfc9095, rfc9108, rfc9338, rfc9683,
rfc9700, rfc9788, rfc9953 round-trip failures

fix: Oga adapter CDATA serialization in plan-based path

The build_moxml_node method always created text nodes, ignoring the
cdata flag on XmlElement. This caused 5 pre-existing OgaAdapter test
failures in cdata_spec.rb on GHA.

fix: Oga adapter CDATA in mixed content and xmlns deduplication

• Check xml_element.cdata flag when creating String child nodes in
  build_moxml_node, creating CDATA sections instead of text nodes
  when cdata is true (matching Nokogiri adapter behavior)
• Filter xmlns attributes in regular attribute iteration to prevent
  duplicate namespace declarations, matching Nokogiri's logic
• Fixes pre-existing CI failures with Canon 0.2.x