Reduce cache-to-execution latency by using lazy unpickling of instructions.

In order to this, invoker generation needs to be cached.

Performance impact on a warm-cache run of test_fmm.py from sumpy

Before

real 43.77
user 44.58
sys 2.96

After

real 37.08
user 37.73
sys 3.02

Closes #67 (closed)

Depends on pytools!12 (merged)