split &fn into &fn and &mut fn

[thestinger]

An &fn would be copyable, because the environment would be immutable. The limitations would match the ones placed on regular borrowed pointers (&mut is moveable, & is immutable) so I think it would result in a more intuitive system.

This would be required to make #8622 (adding dynamically sized bare function objects) backwards compatible but can be implemented independently and will not be an enormous task. All existing &fn would become &mut fn and it will remain the default since &mut coerces to & by-move.

[thestinger]

Nominating for the backwards compatible milestone.

[nikomatsakis]

Current plan, I believe, is roughly what I described here: http://smallcultfollowing.com/babysteps/blog/2013/05/30/removing-procs/. In this case, the only remaining closure type would be fn(...) (note the lack of any sort of ampersand).

[thestinger]

This won't actually work as is (discussed in the linked bug) so I'm just closing it for now.