Rework service discovery to use less memory

[Hexxeh]

Right now the way I implemented this, we build a complete list of services, characteristics and descriptors before we start converting anything to the Pebble structs and notifying the OS.

We could be more efficient about this and do depth-first discovery rather than discovering all characteristics and then all descriptors, instead doing this per service and freeing the memory for a given service once we're done.