More Cartesian housekeeping, plus implementation of copy!

[timholy]

• Cleans up @ngenerate to add return-type annotation
• Adds @nsplat to generate specialized versions of functions that avoid splatting
• Adds quite a few new tests of type inference
• Implements cartesian copy! for faster performance on SubArrays

The addition of @nsplat is not strictly necessary; I could ditch that and write them out by hand, if desired. The main motivation for adding it was to automatically synchronize the use of splatting with changes in CARTESIAN_DIMS.

Performance gain on copy! is the usual:
Master:

julia> A = rand(1000,1000,50);

julia> B = zeros(1000,1000,50);

julia> S = sub(B, :, :, :);

julia> @time copy!(S, A);
elapsed time: 2.144633973 seconds (48 bytes allocated)

This branch:

julia> @time copy!(S, A);
elapsed time: 0.399700238 seconds (48 bytes allocated)

CC @carlobaldassi.

[carlobaldassi]

Good! I'll rebase my branch as soon as this is in.

I thought this change would have allowed getting rid of things like these:

checksize(A, I) = (_checksize(A, I); return nothing)
checksize(A, I, J) = (_checksize(A, I, J); return nothing)
checksize(A, I...) = (_checksize(A, I...); return nothing)

[timholy]

It does, I just missed them---they were just hard to see, since they were in the top 6 lines of the file 😄. Thanks! (If you see any more, don't hesitate to ask, or to fix when you submit your changes.)

Fixed and rebased. Hmm, the original commits are still showing up on this page, but my git history for this branch doesn't show them---anyone know what's going on? Does github now keep copies of old commits when you do a forced update?

[timholy]

Ah, now magically github seems to have gotten this straightened out.

[timholy]

All yours, @carlobaldassi!