fix #6466, deprecate write on non-isbits arrays

[JeffBezanson]

No description provided.

[tkelman]

How will these change when this is removed? The depwarn and invoke calls replaced by errors?

[JeffBezanson]

Yes.

[vtjnash]

This feels a bit wrong, since I think in most cases we've found that dispatching on eltype is annoying / throwing unnecessary errors is painful. In particular, we've generally tended towards f(Any[x, y]) being the same as f(T[x, y]) for all T, and letting the homogeneity of the elements themselves (or lack thereof) define the behavior. But I realize the current behavior isn't entirely consistent either (sometimes opting for direct unsafe_write copies over the more pedantic dispatch loop over all elements), so I'm not strongly opposed either.

[JeffBezanson]

Yes, it's slightly unsatisfying. But the concept is that write is only valid on objects with a "canonical binary representation", and that the direct argument to write (not values inside it) should determine the data format. Otherwise we're stuck with the dispatch loop for all but a fixed set of known types.

Come to think of it the method write(s::IO, x::Ref{T}) where {T} also needs this check; currently write(io, Ref((:a,:b))) writes garbage.