Speed up kernel comparison / loading from cache

Things working:

• !139 (merged)

Things left to do:

• Instructions should not have to be analyzed at all along the fast path from "loading from cache" to "execute".

  • Cache generation of invokers (this is also a non-trivial latency penalty). Invoker generation looks at the instructions and so prevents us from being fully lazy.
  • Kernel.copy() iterates through all instructions; add an option to disable this

• Full use of lazy data structures

  • Add code to generate eq keys and persistent hash keys for instructions (among other things, this has to handle pymbolic expressions, and to normalize the order of sets)
  • Use LazilyUnpicklingList for the list of instructions

cc: @inducer

See also: pytential#38