This adds mutex and block profiles, and brings the command up-to-par
with 'collect-profiles.sh', so that we can remove it.

Profiles are also now collected concurrently, which improves the
runtime from (profile_time * num_profiles) to just (profile_time).

Note that this has a backwards-incompatible change, removing
--cpu-profile-time in favor of the more general --profile-time, which
covers all sampling profiles.