Improve benchmark dataset

[Arwalk]

Looking good, thanks for the iteration, maybe in a future PR we can revisit generating the benchmark dataset using constant sizes for arrays

Originally posted by @menduz in https://github.com/Arwalk/zig-protobuf/issues/135#issuecomment-3356869421

There might be a bit too much randomness into the current benchmark data generator. Limit it for more consistent results, while avoiding situations where we overoptimize certain aspects

Also add some different data types (plain messages with no submessages for example) to vary it.