SNR calculation is very slow, makes continuous mode appear delayed

[zacsimile]

As the title states, the SNR processing makes the data thread take a much longer time to get through displaying all of the images. This leads to non-real-time feedback.

[AdvancedImagingUTSW]

Not a whole lot that can be done here since the conversion is so basic.

S = image.astype(float) - offset_map.astype(float)
    S[S < 0] = 0  # clip
    N = np.sqrt(S + variance_map.astype(float) + 1.0) 

Perhaps the offset_map and variance_map should be converted to type float upon initial loading in the update_snr command? This way it is done once, as opposed to every time? Not sure how much that will shave off the calculation time.

[zacsimile]

@annie-xd-wang more or less circumvented this problem in #533. Converting everything to float on initial loading may help. The sqrt() is likely slowing us down a lot. It's kind of an expensive calculation. Will have a think.

[AdvancedImagingUTSW]

Quickly tried some other ways to do the square root operation, but numpy is pretty damn good. 22x faster than the built in python method for a 2.2k x 2.2k 16-bit image.

Python Version - Duration: 0.044397497177124025
Numpy Version - Duration: 0.002041006088256836