should `add!!(x, t::InplacableThunk)` make sure that x is mutable before calling `t.add!`

[oxinabox]

Came up in JuliaDiff/ChainRules.jl#287

I suspect add!! should take responsiblity as to if it is appropriate to call the InplaceableThunk's add! field.

Its fairly easy for us to setup a baseline set of mutable types:

"returns true if `x .= y` will work to mutate `x`"
is_broadcast_mutable{::Array}} = true
is_broadcast_mutable{::SparseVector}} = true
is_broadcast_mutable{::SparseMatrixCSC}} = true
is_broadcast_mutable{::BitArray} = true  # probably don't need this one
function is_broadcast_mutable(x::AbastractArray)
    p = parent(x)
    p === x && return false  # did not find a mutable parent
    return is_broadcast_mutable(p)
end

is_broadcast_mutable(::Any) = false

ArrayInterface has something like this, but gets it wrong (JuliaArrays/ArrayInterface.jl#77)
and we don't want the dependency

Other types can specialize add!! themselves; with what ever way they like to be mutated.

[oxinabox]

had a chat with @nickrobinson251 on this.
We should do this:

• It makes the interface of InplaceThunk.add! clearer: it only has to handle mutating cases
• It avoids adding lots of functions like _subtract!! into ChainRules etc add rule for matrix negation ChainRules.jl#287, that each individually handle this
• We already kind of do this for the case of the first argument being Zero() anyway.
• Downside is it is a bit more magical, in the way that add!! is already a bit magical: it decides behind the scenes if to be inplace or not
• We should also add a thing in debug mode that checked if the InplaceThunk.add!! did infact mutate.