Return non-zero status code if there are any failures

[astrofrog]

For the astropy-benchmarks, I would like the Travis status to be a failure if there are any failures during the benchmarks (currently there are three benchmark failures but Travis passes). Maybe a flag could be added to this effect?

[bjodah]

For sympy_benchmarks, we would also be interested in a asv --strict flag or something along those lines. But in that case we would also like to be able to mark some legacy benchmarks as allowed failures, e.g. along the lines of:

@asv.mark.may_fail
def time_some_old_method():
    ...

[pitrou]

This would indeed be nice to ensure that benchmarks don't get stale because of changes in program code.