Replace the quadradic duplicate check logic with a linear version

[Stebalien]

It's ~2% slower with 10 hashes and ~2% faster with 100 hashes (with very noisy benchmarks). But the important part is that the runtime should scale linearly with the number of inputs. E.g., this version is 2.5x faster for 1000 signatures.

fixes #68

[Stebalien]

Based on #67.