Clean FLOW count data to improve permanent count station accuracy

[cczhu]

Based on some circumstantial evidence, such as

• The presence of extremely high or low PTC year-on-year growth rates
• AADTs clearly far too low for their road class (eg. 8540609 has only several thousand vehicles per day, rather than several tens of thousands)
• A list provided by Arman of permanent stations that reduce the overall model's predictive accuracy

we know that there are issues in the FLOW data being we're ingesting into the model. These issues can both be categorical errors (eg. accidentally including bike counts), sensor failure, special events, or not properly accounting for locations with major construction or detours. We should clean the data of these issues.

This cleaning process should involve:

• Checking the source code for prj_volume.uoft_centreline_volumes_output to make sure it's only taking in vehicle counts.
• Creating an automated anomaly detection algorithm using Facebook Prophet to spot sensor failure and special events creating unusual volumes.
• Manual follow-up of permanent count locations with unusual growth rates to confirm they should be included.

[cczhu]

Note that right now we use a median to calculate the annual average version D_ijk because PTCs have not been cleaned of high noise locations (we don't for DoM_ijk). get_annually_averaged_ratios should be revisited when this issue is being solved.