Add some latency nofib benchmarks

[ezyang]

At the moment, GHC is heavily throughput oriented (e.g. scheduler choices, etc). However, applications like Cloud Haskell often would like to reduce latency, even if it is at the cost of some throughput. However, we have no good way of measuring when changes to the RTS improve latency. We should add some benchmark programs along these lines.

[igloo]

Thanks for the suggestion. However, I don't think anyone's likely to write such a test based on a ticket being open; they're much more likely to do so while working on improving latency. I'll therefore close this ticket, but of course more benchmarks are most welcome!