[FEAT] Function decorators to specify a conversion policy

This is a follow up on #2813.
Currently, pybind11 provides the possibility to define a return_value_policy for converting C++ objects to Python when returning from a C++ function. However, when passing objects from C++ to Python via arguments of a Python function object, the conversion policy cannot be easily specified. Particularly, this applies to these important situations:

• trampoline methods utilizing PYBIND11_OVERRIDE* as in #2915
• callback function objects as in #1200

In this case, the function's arguments are converted with the hard-coded policy = automatic_reference:
https://github.com/pybind/pybind11/blob/417067eeb8d93e8ed49f3ff580712ae2b9a85259/include/pybind11/pytypes.h#L108-L109
which resolves to reference for pointers and to copy for references. While it was argued several times (e.g. https://github.com/pybind/pybind11/pull/1240#issuecomment-355873711) that pybind11 follows a safe copy strategy by default with the option for users to opt-in for reference passing, this is not easily possible for the above-mentioned use cases.

If one needs passing by reference, one strategy is to explicitly turn reference arguments into pointer arguments as here.
However, I propose to decorate the Python function object:

@pybind11.return_value_policy.reference
def callback(MyObject o):
    o.value = 42

What do you think of this idea, @wjakob, @rwgk, @EricCousineau-TRI, @YannickJadoul?

Start with the conversion call in include/pybind11/pytypes.h at the linked location and the PYBIND11_OVERRIDE* implementation in include/pybind11/pybind11.h. Review follow-up issue #2813 and the callback context from #1200, then determine how the proposed Python decorator should select the argument conversion policy for trampoline and callback functions.