Align benchmark::State to a cacheline.

[ckennelly]

This can avoid interference with neighboring objects and stabilize benchmark results.

[oontvoo]

@dominichamon Another reason to have absl dependency :) Would make this PR so much simpler.

[dmah42]

do you have the results from a run with and without this change?

[ckennelly]

@dominichamon I only have an internal microbenchmark, which demonstrated higher variability across runs (especially when combined with a change to TCMalloc) without the alignment attribute.

[LebedevRI]

This seems plausible (although hard to tell without a test), but is this only about the State? What about ThreadTimer, ThreadManager, PerfCountersMeasurement?