The pool was prone to deadlocks, especially due to having a single pool
for rx+tx. If the pool got full with rx'd packets it would deadlock because
processing a rx packet requires doing another allocation on the pool, for
the possibly tx'd response, before deallocating the rx'd packet.

This also allows Device impls to allocate the packet memory in a particular
RAM kind, if needed for example to do DMA.

The `Device` trait is now token-based, like smoltcp's. In the end, this
is better because it allows callers to manage memory however they want (including
with a pool if they want to).

The "must not be called reentrantly" invariant is too "global" to
maintain comfortably, and the cost of the RefCell is negligible,
so this was a case of premature optimization.