Added a faster linear interpolation scaler to IMB_scaleImBuf.
Only handles common cases when we either
enlarge both, width and height or
shrink both, width and height
but that is pretty fast:
* does only blit once instead of two passes like the old code
(fewer cache misses)
* uses fixed point integer arithmetic for byte buffers
* doesn't branch in tight loops
Should be comparable in speed to the ImBuf ..._fast functions at least
for byte-buffers.
For the rather seldom situations that e.g. width is enlarged and height
is shrunken, the old code is still in use.
-------
Also small fix in sequencer: free float buffers on read now.
19e37e4a35