Add `season_length` argument to `TimeSeasonality`

[jessegrabowski]

I saw this feature in this tfp statespace example.

When you have data with multiple frequencies, this argument would allow you to "stretch" the effect over multiple time steps. Currently, we assume that each time step of the data is one seasonal change, so if you have hourly data and want to do something like observed =hourly_effect + day_of_week_effect, the hourly effect currently works (with season_length = 24), but day_of_week effect (with season_length=7) wouldn't do what we want.

I want to be able to specify that a day is 24 timesteps, which would result in an effect that looks like this (from the TFP notebook):

Notice how the effect is flat for 7 periods, then shifts to the next day-of-week effect.

[gdeleva]

@jessegrabowski Thanks a lot for your guidance. I opened a draft PR #509 that includes the definition of seasonal component with the new parameter num_steps_per_season. If you have time, could you please confirm that this is exactly what you mentioned above before I start the actual implementation?