Reintroduce single cache for successive scanline reads

[kdt3rd]

Previously we allowed multiple threads to cache but that caused memory bloat, here, only keep one so successive reads will be faster for compression types where there are more than one scanline per chunk

Fixes #1887

[haaninjo]

Thanks, this greatly helps opening speed in GIMP.

Tested and clocked opening 3 exr images created in Darktable:

File	Opening time before	Opening time with this patch
A	6s	Instant
B	32s	<2s
C	65s	<2s

[cary-ilm]

Looks good to me, except does it break the ABI?

abipkgdiff reports that it detects data members that have changed sizes.

[lgritz]

Looks good to me, except does it break the ABI?

abipkgdiff reports that it detects data members that have changed sizes.

I think this might be a flawed process for detecting ABI changes. This patch doesn't change any public headers. The class whose size is changing is totally internal.

There ought to either be an explicit list of these classes and functions to exclude from ABI checks, or better yet, enclose all of them in a private namespace so they can be excluded in one fell swoop.

[kdt3rd]

The public ABI does not change with this, as this is in the pImpl internal class. However, is declared as a private struct within the public class such that there can be a non-void pointer in the public class. We could make it an empty struct type and derive a truly private subclass in a namespace, but would add a bunch of casts or double pointer for little gain?

Detecting ABI changes is a bit fraught, and so I think at best it should just be reviewed manually and not something to be relied upon automated testing for - a simple equal-size typedef change can trigger an apparent abi difference, even though not actually any different.

[kdt3rd]

for the next release (3.4 or 4.0), I would like to take another review of public / private symbols as there are a lot of "utility" classes exposed that should not really be part of the public interface (although have been, so will need to go through a deprecation period), perhaps there is a different way we could achieve this such that abi diff tools can better ignore (i.e. making the pImpl class as a hidden class might work?)