Improve AGC

[bgottula]

A crude AGC was added to the capture software under #1 but it was not extensively tested. Known or suspected issues:

• Slow initial adaptation when the gain/exposure time is far from where it needs to go (several seconds).
• The mapping of agc_value to gain and exposure time settings needs to be further scrutinized.
  • Does the pixel value change linearly in response to a change in gain? What about exposure time?
  • Right now, the mapping minimizes gain to maximize SNR. But we may also want to keep the exposure time low to reduce motion blur induced by mount vibration and tracking jitter.
• More thought should be applied to how the error term is derived and the set point.
• With the lens cap on the AGC does not always crank the gain and exposure time all the way up to max, which means it must be detecting pixels with very high values. But I don't see any non-zero bars in the histogram display. (Unfortunately the histogram used for AGC is generated separately from the one shown in the preview thread, so there could be a bug in one or the other or both.)

[bgottula]

An idea tried on nix was to set the histogram threshold based on a number of allowed saturated pixels, rather than a percentile:

         // Calculate Nth percentile pixel value
-        constexpr float percentile = 1.0;
-        uint32_t integral_threshold = (uint32_t)((1.0 - percentile) * Frame::IMAGE_SIZE_BYTES);
+        constexpr float percentile = 0.999;
+       constexpr int max_saturated_pixels = 10;
+        //uint32_t integral_threshold = (uint32_t)((1.0 - percentile) * Frame::IMAGE_SIZE_BYTES);
+        uint32_t integral_threshold = max_saturated_pixels;