Remove PathBias use of timestamp_dirty and record usage accounting
Instead of using timestamp_dirty as a signal for "circuit use" in the pathbias code, we should create a dedicated path_state_t state (PATH_STATE_USE_ATTEMPTED).
It's also debatable as to if we should remove the additional timestamp_dirty markings I added in legacy/trac#7157 (moved). The most controversial addition was marking freshly cannibalized circuits as dirty. While this seems to make sense to me (do we really want to use cannibalized circuits for any random stream?), it may have subtle impacts on the proper use of GENERAL purpose circuits that get cannibalized during predictive circuit building or other cases involving hidden services.