don't loop on too-flat CGM data

[scottleibrand]

As detailed at #1257, a failing Libre sensor, uploaded to Nightscout via LimiTTer/xDrip+, can result in falsely high and too-flat CGM readings and incorrect insulin delivery. The existing mitigation for this, to avoid looping on completely unchanging CGM data, was insufficient in this case. To further mitigate this and similar issues, this PR changes that check to look for CGM data that is changing less than 1 mg/dL/5m for 45m. As shown at https://gist.github.com/scottleibrand/4c8c84d9989afc98bc92214a08e849f0, that would have been sufficient to prevent further insulin dosing in this case after CGM data falsely flattened out at 290 mg/dL.

This will likely result in occasional false-positives when BG is truly flat for an extended period of time, or transiently when both the short_avgdelta and long_avgdelta are between -1 and 1 mg/dL/5m. However, such false-positives should be brief, and will simply result in OpenAPS refraining from SMBs or high-temps until more-plausible CGM data arrives.

[scottleibrand]

Fixed tests. Have been running without issue on our rigs for over a week now. Ready to merge if no objections.

[scottleibrand]

Also reviewed all the logs from the rigs this was running on, and confirmed that the largest microbolus delivered after a period of too-flat CGM data was 0.1U. Also updated the too-flat logic to only kick in when BG > 60, to avoid the possibility of a zero-temp expiring prematurely when BG is low and flat.

[scottleibrand]

Since 5962202 this is now looking good on our rigs. Ready to merge if no objections.