Apply restrict optimizations to ImagingQuantize's output loop

[akx]

Follows up on #9752 (comment) and #9740.

None of the other loops in the quantization "frontend" were eligible for restrict, but this does speed up quantization a little bit, more so when the quantization method itself is fast (such as FastOctree).

Note the below local benchmark is with larger images than the default. CodSpeed probably won't show comparisons here since the test IDs change.

------------------------ benchmark 'quantize': 5 tests, 2 sources -----------------------
Name (time in ms)                            0001_5b0d565 Min  0002_b945920 Min      ΔMin
-----------------------------------------------------------------------------------------
test_quantize[2474x1622-RGBA-FASTOCTREE]              12.9134           11.4034    -11.7%
test_quantize[2474x1622-L-MAXCOVERAGE]                79.4019           77.5946     -2.3%
test_quantize[2474x1622-L-MEDIANCUT]                 141.6351          140.1765     -1.0%
test_quantize[2474x1622-RGB-MEDIANCUT]               392.6901          382.9610     -2.5%
test_quantize[2474x1622-RGB-MAXCOVERAGE]           4,097.8388        4,015.4798     -2.0%
-----------------------------------------------------------------------------------------

I also noticed MAXCOVERAGE is painfully slow with images with a large number of distinct colors, but I think that would require a more thorough rework. By and large, the quantization code (and its hash table code) is from 1998 or so, written with machines from that era in mind.