Make untyped mutating callbacks work

[tkf]

I'm writing this PR to start a discussion: Do we want untyped Julia callback to be able to mutate Python objects? That is to say, do we want the following code to work?

py"""
def apply(f, *args):
    f(*args)
    return args
"""
@test py"apply"(fill!, zeros(3), 10)[1] == [10, 10, 10]

I got here while trying to fix pyjulia bug I observed in JuliaPy/pyjulia#183 (comment)

It requires some run-time introspection and hard-coding many mutable Python objects (I only did it for Numpy array at the moment). Is there a better way to do it? Do we want this in the first place?

[tkf]

Maybe it makes sense to move the code above inside convert(::Type{PyAny}, o::PyObject)?

[tkf]

Maybe I should have called it apply_inplace or sth to avoid confusion with Python 2 builtin apply.

[stevengj]

Julia callbacks mutating Python objects will only work (in general) if you use pyfunction to control how the Python arguments are converted for the Julia caller.

For example:

julia> pyfill! = pyfunction(fill!, PyVector, PyAny)
PyObject <PyCall.jlwrap PyCall.FuncWrapper{Tuple{PyVector,PyAny},typeof(fill!)}(fill!, Dict{Symbol,Any}())>

julia> v = py"[1,2,3,4]"o
PyObject [1, 2, 3, 4]

julia> pyfill!(v, 7)
4-element Array{Int64,1}:
 7
 7
 7
 7

julia> v
PyObject [7, 7, 7, 7]

I think this is the right way to do it — for example, you can't rely on everything being a NumPy array.