Fix restricted variants to try multiple candidates

[beckjake]

Currently, if you define multiple variants that meet the same explicitly restricted fields, dbt will just try the first one and fail if it can't parse. This can have counter-intuitive consequences:

@dataclass
class ParsedModel(JsonSchemaMixin):
  resource_type: NodeType.Model = field(metadata=dict(restrict=[NodeType.Model]))

@dataclass
class CompiledModel(JsonSchemaMixin):
  resource_type: NodeType.Model = field(metadata=dict(restrict=[NodeType.Model]))
  compiled: True

@dataclass
class UnionedModel(JsonSchemaMixin):
    value: Union[ParsedModel, CompiledModel]

# good!
assert UnionedModel(ParsedModel(NodeType.Model)).to_dict() == \
    {'value': {'resource_type': 'model'}} 
# good!
assert UnionedModel.from_dict({'value': {'resource_type': 'model'}}) == \
    UnionedModel(ParsedModel(NodeType.Model))
# good!
assert UnionedModel(CompiledModel(NodeType.Model, True)).to_dict() == \
    {'value': {'resource_type': 'model', 'compiled': True}}

# horrible error, don't even need to assert!
# "'compiled' was unexpected" - only tried to parse as a ParsedModel
UnionedModel.from_dict({'value': {'resource_type': 'model', 'compiled': True}}) == \
    UnionedModel(CompiledModel(NodeType.Model, True))

Reversing the order of the union will reverse the issue.

This PR fixes that issue by trying each variant. I refactored the restricted/unrestricted logic a tiny bit since we're now handling them the same way for both restricted and unrestricted fields.

An alternative I considered was making a special restriction that says "this field must exist". That required a bunch of changes to validating and parsing restrictions, while this seemed fairly straightforward.