fix pin violation in much of concurrentstream

[conradludgate]

Fixes #182

I developed the crate futures-buffered a couple years ago with the intent of being a more efficient 'FuturesUnordered' type than the one in futures_util.

The core of the unsoundness in #182 are the pin-violations caused by FutureGroup slab reallocs. futures-buffered solves this by using a 'triangular array' pattern, where each subsequent realloc is actually a new slab.

Note: the unsoundness still exists in FutureGroup. futures-buffered does not return keys for entries (although that is not a bad idea) so the API is not compatible.

A quick soundness fix for FutureGroup would be to make Stream require F: Unpin, albeit this is a breaking change.

Even if you do not merge this change, I hope it serves as a nice inspiration for you :)

[yoshuawuyts]

Overall this looks great, thank you so much! - I'd like to eventually update our own {Future,Stream}Group types to not have the issues. And actually would prefer to keep external deps to a minimum.

But I'm also recognizing that we shouldn't let perfect be the enemy of good, and this actually fixes our issues straight away - so I'm keen to accept this patch and then figure out where we go after that.

If you could revert deleting the methods in future_group, I think we should be good to merge! Thank you so much!