Constraints on ID size

A lot of messages contain strings that are (typically) 32 bytes hexadecimal strings.

As I am working on a native implementation of the protocol, I realize that a lot of optimization is possible in e.g. comms if we know the size of the id in advance. Basically, a comm contains is a string, and fixed-sized stack allocated strings would prevent requiring dynamic allocation when instantiating a comm.

Hence, after some discussions on gitter, a proposition is to require the IDs specified in the jupyter messaging protocol to have a length of *up to* `32 bytes`.