Processing ByteStream data in python is full of performance traps.
Juggling python and C++ types together to decode HLT EDM collection
names from a BS file makes it even worse. Optimise the script by
reducing the most costly operations, mainly iterations over raw data,
by reworking loops and functions, and using lru_cache to avoid
decoding the same words multiple times.

Cuts down the time it takes to process 1000 events by a factor of 2
from 400 to 200 seconds.