Benchmarks: use context pooling (almost) everywhere

[roji]

All our benchmarks should be using context pooling, to avoid the overhead of context creation interfering with results. We should still keep one benchmark to continue measuring pooling vs. non-pooling perf.