Improve efficiency of HysteresisDeadband::is_in_deadband

[rhaschke]

The average over last errors, this class uses a std::deque to maintain a sliding window of previous errors:

sr_core/sr_utilities/include/sr_utilities/sr_deadband.hpp

Line 90 in 9d90e75

last_errors.push_back(error);

sr_core/sr_utilities/include/sr_utilities/sr_deadband.hpp

Line 136 in 9d90e75

last_errors.pop_front();

While this approach is semantically correct, it is highly inefficient, copying 50 values around in each call.
There are much more efficient ways to just a compute a sliding average:

• Use a circular buffer, maintain the average sum, and just subtract the popped and add the pushed value (all others values remain the same).
• Use estimated sliding average (which should be sufficient here):
  avg_error = (1.0 - lambda) * avg_error + lambda * error
  where lambda adjusts the smoothing and should be chosen lambda=1.0/nb_errors_for_avg here.

I strongly suggest the latter method and storing lambda once in the class (nb_errors_for_avg shouldn't change over time anyway!) As this requires API changes, I leave it to you to decide on this and implement it.

[toliver]

I agree we should make this more efficient. We've also found recently that the default average of 50 values introduces an undesirable zeroing of the error ~25ms after a zero crossing.
We should have expected this, of course. The default averaging window should be much narrower (or not use averaging at all) to get rid of this delay.