Refine HateoasHints to generate a more focused reflection configuration

[sdeleuze]

The analysis that has ended up in creating this issue started from a com.oracle.graal.pointsto.util.AnalysisError: guarantee failed error on https://github.com/odrotbohm/spring-restbucks, which has shown that the reflect-config.json entry responsible for this error was:

{
    "name": "java.lang.Class",
    "allDeclaredFields": true,
    "allDeclaredConstructors": true,
    "allPublicConstructors": true,
    "allDeclaredMethods": true,
    "allPublicMethods": true,
    "allDeclaredClasses": true
}

After a deeper look, it appears that HateoasHints is generating far too much reflection entries via (at least) HateoasHints#computeRepresentationModels and HateoasHints#computeAtConfigurationClasses.

HateoasHints#computeRepresentationModels should probably be refined in order to at least filter out java.* types or to only create a { "name": "java.lang.Class" } without all the other flags.

HateoasHints#computeAtConfigurationClasses is suspicious, do we really need such inference with the new AOT engine? It should IMO be removed or restricted to a much more limited set of types.

Last point (but likely first one to implement), it would be great to be able to reproduce this issue on the hateoas sample, if that can help one of the guilty classes seems to be javax.money.CurrencyUnit which trigger the analysis of javax.money.AbstractContextBuilder which ends up with the problematic java.lang.Class entry.