Generate _out functions from native function specifications

[gchanan]

[gchanan]

Here's a sketch of how this could work for a new native function foo with a Tensor arg self:

1. User writes the "foo_out" variant as a native function, e.g.: at::native::foo_out(Tensor &result, const Tensor &self); and marks result as an output: True argument.

2. In base Type, we generate foo(self) and foo_out(result, self). foo_out just calls the native function, e.g. at::native::foo_out(result, self). foo generates a result Tensor of the correct type and passes it on to foo_out. I'm not sure if this result Tensor generation works today in the base type or we have to do the generation in the derived type; the current _out variants are generated in the derived type, but it's unclear to me if that's necessary or just because the non-out variants are also generated in the derived type (and they share code) so it's the most straightforward.

3. In Function, we generate foo and foo_out that call the corresponding Type variants after inferring the type.

4. In Method, we only generate foo that calls the Type variant foo foo

[gchanan]

BTW, I'm not sure we want to implement this fully yet; Variables don't currently work with the _out variants (none of ATen-backed, C++ pure, python pure autograd functions), so we probably want to figure out a story there before implementing this fully.