Go-faster capability

[sageserpent-open]

Writing a benchmark that uses Americium solely as a generator of data inputs in the benchmark taking a lot of time and memory purely in constructing the inputs, irrespective of the actual code whose performance is being measured.

Using a simpler approach of hand-rolling some random number generation and set construction yields much better performance, see the updated benchmark without Americium.

Either come up with some optimisations to improve Americium's performance when supplying a very large number of cases, or provide a configuration option that turns off some of the bookkeeping, sacrificing, say, the uniqueness guarantee or the ability to reproduce or shrink a failing test case. The latter two aren't important for benchmarking, so if they are causing a lot of overhead, disable them.