Migrate 2D image cache away from std::vector<float>

[veggiesaurus]

We currently use a std::vector<float> for Frame::_image_cache. However, the ability to resize the image cache is not of any use, because a new cache is made for each separate image, and all image channels have the same 2D slice size. The main problem with using a std::vector is that we end up allocating memory and initialising to zero before we fill it with values from disk. This means it will always be slower than something like a simple c-style array.

This is not a problem for small images, but for large images it can become a substantial portion of the initial image load time.

[jolopezl]

@veggiesaurus, do we have any design constraint to resist introducing c-style arrays here in favor of a significant performance gain for large images?

[veggiesaurus]

No, thats totally fine, as long as they are carefully cleaned up. We don't actually ever need to resize the image cache if I recall correctly