Method `__getattribute__()` is documented as the "way to to actually get
total control over attribute access" [1] so we change the implementation
of slotted classes with cached properties by defining a
`__getattribute__()` method instead of `__getattr__()` previously.

[1]: https://docs.python.org/3/reference/datamodel.html#customizing-attribute-access

Just changing that preserves the current behaviour, according to the
test suite, but also makes sub-classing work better, e.g. when the
subclass is not an attr-class and also defines a custom __getattr__() as
evidenced in added test.

In tests, we replace most custom `__getattr__()` definitions by
equivalent `__getattribute__()` ones, except in regression tests where
`__getattr__()` is explicitly involved. Also, in
test_slots_with_multiple_cached_property_subclasses_works(), we replace
the `if hasattr(super(), "__getattr__"):` by a `try:`/`except
AttributeError:` as using `hasattr(..., "__getattribute__")` would be
meaningless since `__getattribute__()` is always defined.

Fix python-attrs#1288