make xml reader and writer extensible

[tomjnixon]

Use cases:

• read or write metadata from a newer ADM standard, or during development of new standards
• read or write custom metadata that extends existing ADM structures e.g. custom parameters
• read or write custom metadata that lives in AXML but outside of existing structures, e.g. tool version information

Implementation ideas:

1. extend all ADM elements with xml attributes, and use it to (optionally) store a reference to the corresponding XML elements when parsing or generating XML
   • when reading, this can be used to extract extra data from the xml
   • when writing, extra data can be injected into the XML before it's string-ified
   • bad: feels like writing an ADM document shouldn't modify it
   • bad: slightly increases storage even when not used
2. allow customisation of the XML to python class mapping
   • need to be able to tell parser "when dealing with element X, use this parser and this class"
   • need to be able to tell generator "when dealing with this class, use this un-parser"
   • bad: exposes innards of XML parser to users
   • bad: quite complex to do, especially for block formats. XML parser is already complex enough
3. when parsing or generating XML, produce a mapping from ADM objects to XML elements
   • question: how to store this mapping?
     • {id(adm_element): xml_element}
     • {id(adm_element): (adm_element, xml_element)}
     • {IDCompare(adm_element): xml_element} where IDCompare is a class that makes an object hashable and comparable by identity
     • some hack to make WeakKeyDictionary work with non-hashable types. It's not clear if weak references are really needed
   • could pass a list into the parser / generator to say which types we want to store
   • could be combined with 2, allowing the user to specify their own sub-classes, but keep parsing / generation external. It should be possible to re-use the parsing infrastructure externally, but this still allows the internal structure to be modified.

After writing this I think 3 is the way to go.