builds() and from_type() should introspect attrs-classes

[Zac-HD]

There's heaps of metadata on the attributes, and we could certainly infer a strategy for attributes with an "instance_of" validator. Even with python-attrs/attrs#249, special handling would still be valuable because we could filter by any other validators.

This issue brought to you by realizing that I wanted this feature in a project I'm playing with - in that case it's going to be faster to hard-code a strategy, but it would definitely be used as well as nice to have 😄

[CodyKochmann]

Ive had to reroute a few projects because I was using that type of layout but didn't have a clean way to test specifically this so definitely +1

[Zac-HD]

Some notes:

• attrs supports extensions by providing Attribute objects.
• attr.fields(SomeClass) returns a tuple of Attributes for SomeClass
• An Attribute may have a default of a value or a factory; and may also have a type. These are easy to construct a strategy for.
• In the unlikely event that default==NOTHING and type is None but convert or converter (at most one may be non-None) is a callable with type annotations, we may be able to infer a strategy that way.
• Converters may also be wrapped in converters.optional, which makes None a valid value.

This all becomes more "interesting" once we start to consider validators.

• Built-in validators are great for introspection. optional as for converters. An instance_of validator is an older way of specifying type. An in_ validator is a perfect feeder for sampled_from.
• Multiple validators can be given as a list or wrapped in and_; each are run in order so we'd need to unwrap them.
• Validators run on the state-so-far, and may depend on multiple fields, so we'll need to have some logic that converts validation errors to a reject() call (as is done for Django models)

All together, this probably means we'll write a new module hypothesis.searchstrategy.attrs and function from_attrs_class, then if isinstance(thing, type) and attr.has(thing) return from_attrs_class(thing, *args, **kwargs) - as for builds itself and e.g. Django models, we want to allow users to set sub-strategies explicitly - at the top of builds (and from_type).