Use correct min/max stretching, and guard against flat images

[thouis]

The min/max stretching was using (im - min) / max, instead of (im - min) / (max - min). This change fixes that stretch, and also adds a small epsilon to guard against NaNs when max == min.

[thouis]

I also added a check for the image size matching the configured size, and a warning for cell boxes entirely outside the image.

These changes all originated from the same run,
With a misconfigured image size,
That led to clipping boxes outside the image,
Which produced all-zeros images,
Which produced NaNs in the image normalization.

[thouis]

The assertion on image size could be ignored if you first incorporate this PR: #354

[jccaicedo]

Looks good to me!

[jccaicedo]

Thank you for the PR!! This is amazing!