Improper handling of subschema validation in unevaluatedProperties.

[tobz]

As part of #419, I chose to make the subschema subvalidator function on a first-match-found basis, such that if you have a oneOf with multiple subschemas, the first subschema found that would theoretically evaluate the property at all is the result chosen for the overall validate/apply call on the subvalidator.

However, this is wrong, and leads to unintuitive errors in certain cases. For example, if the schema has oneOf where all subschemas are object schemas with a single, identical property, but differing values ("const": "<value>", in my case), then validating that property against any of the subschemas would theoretically result in an application that gave back a result -- because the schema does care about the property -- but may or may not fail depending on the order of the subschemas, and the intended subschema being checked first or not.

In general, we should be obeying the specific behaviors of the given subschema validation mode: continuing with the oneOf example, we should be checking all subschemas, and ensuring that only one of them that indicated that it evaluated the property was in fact valid. Similarly, behavior for allOf and anyOf would need to modeled with their respective constraints.

I'm going to work on a PR to do the aforementioned changes -- treat each subschema validation type (oneOf, etc) distinctly -- along with unit tests, now that I can see that I need to be more deliberate about ensuring that validator conforms to the behavior outlined in the specification. This is, I believe, a consequence of the design I implemented -- checking property evaluation directly rather than collecting annotations at the end, like most third-party validators seem to do -- but I do still think it can fixed without much effort and in a clean and internally consistent way.