Makes all objects in connection have a stable address

[anarthal]

Moves read_buffer memory reservation to the connection's constructor
Makes read_buffer memory reservation size be a power of 2

[anarthal]

This is intended to alleviate one of the problems you had in #283, and make reasoning about lifetimes trivial.

Before I add some tests to verify that moving connections is fine, do you think this is the way to go?

I could also make exec_op, reader_op, writer_op and friends operate on connection_impl rather than connection, which should be marginally more efficient, but requires more work. Let me know if you think it's worth it.

[mzimbres]

AFAICS this is only needed if we want to support a connection being moved while async_run is pending. Or is there anything else? Is it reasonable to allow that?

[anarthal]

It's aimed to solve things like this: #283 (comment)

This would cause a segfault after moving the connection, whether you perform the move before, during or after async_run executes. IMO not having to reason about lifetimes at all makes things much easier.

[mzimbres]

Hi Ruben, I am going to address those comments in the way you suggested. Just don't have much time now. I agree what you propose here is good (and common) practice and does not have any bad side effects. I support further development here, thanks.

[anarthal]

Great. I will do some further fixes here and post it for review again.

I'm going to be out of office during the weekend too, so don't worry.

[anarthal]

This is now ready @mzimbres. I've made all composed operations use connection_impl, and moved some of the functionality required by them there. I've also moved the reserve() call to the connection's constructor.