What is the problem we try to solve when we raise the gain at all? Is it that we raise Gain to reduce exposure time so we have less of a chance of tracking/seeing errors?
If we assume we have perfect tracking and perfect seeing, would Gain 0 (10 offset) be the most optimal?
Looking at the ADU values above for optimal exposures, would all of those exposures be equal to each other or would the 0 gain one be superior and have more dynamic range than the others.
Increasing gain can cost you dynamic range. However, because at a higher gain you reduce read noise, you simply don't need long exposures.
This camera uses a 12-bit ADC, rather than a 16-bit ADC. This means the quantization error is higher. With the average CCD, quantization error is around 0.1-0.3 ADU. With the ASI1600 at Gain 0, it/s 1.4 ADU! To average out this quantization error, you need more subs. Now, at a higher gain, the quantization error actually drops. At Gain 76 it is 0.6 ADU. At Gain 200 it is 0.14 ADU. So higher gains greatly improve the sampling of each electron worth of signal, giving you a more accurate result.
The camera is still 12 bit though, and in practice stacking only a few subs at any gain results in posterization, sudden large jumps in levels, or a "gappy histogram." To smooth out the noise profile, you need to stack more subs. You can recover around 3 bits by stacking 80-120 subs. You can recover around 4 bits by stacking 250-300 subs.
I recently just stacked 413x30s subs, Gain 76/15, with aggressive dithering every 4 subs, and the results were a very nice clean noise profile with ~16.4 bits of real-world precision:

Closeup of the noise profile:

This was done under pretty poor skies, 18.1-18.3mag/sq" during the nights I imaged. I plate solved the field, and plotted objects as faint as 22mag/sq". The faintest in the entire field was over 21mag. The faintest within the crop here is 19.49mag, which is still quite faint considering how bright my skies were:

Lower bit depth CMOS cameras work really well when you dither aggressively and stack a lot of frames. To stack a lot of frames, you want the lowest read noise you could possibly get. At Gain 0, the total read noise from 413 subs would be 71.13e-! At Gain 76, the total read noise is 40.6e-. If I had dark enough skies, I'd probably use unity gain, which would give a total read noise of 31.5e-. For the record...stacking 20 long exposure KAF-8300 frames would give you total read noise of 40.3e-. 