Backport split parser changes

[treeman]

This backports features and fixes from the split parser experiment but keeping them in a single parser.

This is a major rewrite including:

• Complete rewrite of inline handling, aimed to resolve Djot's precedence rules
• Added tree-sitter fields wherever I could, so you can refer to child nodes by names instead of only indexes
• We're doing more work but performance should be roughly the same, requiring some extra logic to prune branches
• And fixes for many different parsing problems

Highlight groups have some breaking changes:

• Removed header1, header2 etc in favor of a header node that can contain an arbitrary level.
• Superscript, subscript, highlighted, insert, and delete now use named nodes like superscript_begin and superscript_end instead of text nodes like ^ and {^.
• Standalone comments such as {% comment %} are now captured by the inline_comment node instead of in a comment node (which is used for comments inside an attribute such as in {.class %comment% #id}).
• image_description now contains the ![ text node for images of all types (previously captured by full_reference_image, collapsed_reference_image, and inline_image)

Closes #41, #42, #43, #44, #45, #46, #49, #50, #52, #53, #54