Add benchmarks

[qmfrederik]

[davidchisnall]

We have a benchmark like that in one of the tests (conditionally compiled, off by default). It's a bit misleading because it's always hitting the same cache line and will always hit in the branch predictor. Last time I did some deep analysis, it was taking around 4 cycles on a Haswell chip to go through the entire message-send machinery. That drops off sharply if we miss in the L1 for the dtable.

Running benchmarks in CI on GitHub Action Runners is not very useful because they'll end up on a random VM (where CPUID lies and may pretend to be an older core than it is), so may vary hugely with no code changes.