Avoid simple try-all deserialization for type choices when possible

[rooooooooob]

e.g.

transaction_metadatum =
    { * transaction_metadatum => transaction_metadatum } ; @name map
  / [ * transaction_metadatum ] ; @name list
  / int
  / bytes .size (0..64)
  / text .size (0..64)

For type choice we just start trying variants and backtrack if we encounter an error. This allows us to simply support all type choices with one codegen method, but has several downsides when the type choice consists of completely different types like the above. One downside is we waste time trying the other variants, but more critically is that errors are essentially eaten completely making debugging more of a pain. Instead of getting specific errors telling us exactly what went wrong and where, it will be eaten up by the topmost type/group choice as simply "No variant matched".

This would complicate codegen slightly but I think it's worth it just for the better errors. I just wasted some time debugging CML block parsing that would have been made immediately obvious if we did this instead. The faster deserialization time for choices would be an added bonus.

It can be solved by matching on raw.cbor_type() and having branches for each covered type in the choice. If the user isn't using extern types we have the ability to check all possible cbor starting tags for types, although the support isn't 100% for structs (or at least plain group ones?) it is better than nothing and would still help us for most cardano CDDL specs.

[rooooooooob]

Another option could be to just give DeserializeFailure::NoVariantMatched a Vec<DeserializeError> value which would require less changes. Or we could do the above match for when they're disjoint, and the old approach with the vec of errs when they're not.