Remove impl Sync for FuturesUnordered

[okready]

FuturesUnordered was previously modified to use interior mutability for
the len and head_all fields using Cell, but its Sync implementation was
left intact. This implementation is no longer valid, so it has been
removed.

Fixes #2050.

[seanmonstar]

I suspect code already depends on this bound, so it may be better to add the synchronization in instead of removing this bound, otherwise people's code will stop compiling in a patch update.

[okready]

Good point. I realize simply wrapping the internals in a Mutex isn't ideal, but FuturesUnordered::push is a bit non-trivial, so it seems like a reasonable choice for now without requiring any major rewrites, and it shouldn't slow down polling or other &mut self methods since they can simply call get_mut on the mutex. Any objections?

[okready]

Disregard that, I didn't realize Mutex (the thread type, not the futures type) is only in std.

[Pauan]

AtomicUsize is available in core (and it should be used anyways, because it's faster and smaller than Mutex).

[okready]

Fixing push to work concurrently using atomics seems straightforward (len and head_all can be briefly out-of-sync without things breaking). There's also the interaction between push and iter though. iter simply grabs the current head_all and len values, but they might not match up if head_all and len aren't kept in sync with each other.

We could just ignore the len field when creating an IterPinRef and either compute the length manually when creating the iterator or lazily when calling size_hint. It could be costly for large FuturesUnordered lists, but it would avoid having to do any additional synchronization between the two fields.