Description of new feature

This issue was originally very long, but the fundamental motivation is quite simple. Right now, behaviour classes cannot perform validation at construction time; the only opportunity for validating the current layout is whenever a user-defined method is called, e.g.:

class OnlyFloats(Array):
    def validate(self):
        dtypes = []
        def get_dtype(layout, **kwargs):
            if isinstance(layout, ak._v2.contents.NumpyArray):
                assert np.issubdtype(layout.dtype, np.floating)
        self.layout.recursively_apply(get_dtype)

ak.behavior["only_floats"] = OnlyFloats

>>> x = ak.with_parameter(Array([1,2,3]), "__array__",  "only_floats")
>>> x.validate()
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
/tmp/ipykernel_362562/672691759.py in <cell line: 1>()
----> 1 validate(ak._v2.Array(l))
...

/tmp/ipykernel_362562/3059306649.py in get_dtype(layout, **kwargs)
      3     def get_dtype(layout, **kwargs):
      4         if isinstance(layout, ak._v2.contents.NumpyArray):
----> 5             assert np.issubdtype(layout.dtype, np.floating)
      6     self.layout.recursively_apply(get_dtype)

AssertionError:

This means that a user could associate a behaviour with a poorly structured array, and only discover their error at a later point in the program when the validation method is called.

I wonder if we should consider adding a hook that allows behaviours to run user-defined code at the time of array wrapping, e.g. to validate the array:

class Array(...):
    def __init__(self, ...):
        self._on_layout_added()

    def _on_layout_added(self):
        pass


class OnlyFloats(Array):
    def _on_layout_added(self):
        dtypes = []
        def get_dtype(layout, **kwargs):
            if isinstance(layout, ak._v2.contents.NumpyArray):
                assert np.issubdtype(layout.dtype, np.floating)
        self.layout.recursively_apply(get_dtype)

ak.behavior["only_floats"] = OnlyFloats


x = ak.with_parameter(Array([1,2,3]), "__array__",  "only_floats")

This would mean that we don't have to come up with some Awkward-specific validation tool: this can be left up to the library authors.

The caveat here is that I don't think we want this validation step to do anything too magical - ideally the mental model of behaviours as a mechanism to attach methods to classes holds.