Speed up getpixel() and get_flattened_data()

[akx]

Came up as a side-effect of #9292 + https://github.com/akx/Pillow/pull/13/changes#r3819140811. Doesn't change the memory price of get_flattened_data() though.

Locally (pay no heed to the benchmarks' commit IDs; this final version has less redundant mode tests and as such, different commits):

---------------------------- benchmark: 8 tests, 2 sources ----------------------------
Name (time in ms)                          0001_03fd7da Min  0002_e70ed66 Min      ΔMin
---------------------------------------------------------------------------------------
test_get_flattened_data[1237x811-I]                  6.8375            6.9130     +1.1%
test_get_flattened_data[1237x811-P]                  6.9563            6.9265     -0.4%
test_get_flattened_data[1237x811-L]                  7.0001            6.9799     -0.3%
test_get_flattened_data[1237x811-1]                  7.0226            6.9455     -1.1%
test_get_flattened_data[1237x811-F]                 15.2092           15.2749     +0.4%
test_get_flattened_data[1237x811-LA]                51.0383           38.8171    -23.9%
test_get_flattened_data[1237x811-RGB]               59.7311           43.9531    -26.4%
test_get_flattened_data[1237x811-RGBA]              66.1110           47.2317    -28.6%

Something that isn't necessarily captured by the benchmark is the PyObject_GC_UnTrack change; since we know that these tuples (immutable, as tuples are) will only contain 8-bit integers (and as an aside, CPython has likely returned cached small-integer references), we can tell the GC to not bother peeking inside them to see if there are cycles. This means that the next GC after get_flattened_data will not need to do that itself (it does need to do it for the tuple() cast from get_flattened_data once though).