Fast image allocation

[homm]

The Pillow's ImagingNew function always creates images filled with zeroes (black color). While this is fine for some operations, for others this is absolutely not required and can affect performance, especially for operations which does not required intensive computation.

Zeroing is required if:

• operation depends on initial pixels values and expects zeroes
• operation can be applied on the part of the image
• operation can be applied partially, for example loading corrupted image

In other cases zeroing can be avoided.

This PR provides new ImagingNewDirty function which makes no warranties about the content of created images. Also, this PR uses ImagingNewDirty for many operations, such as resize, crop.

[homm]

So, there are benchmarks. I've tested preview of Pillow-SIMD 4.3 on desktop i5-4430 on not virtualized Ubuntu 16.04.

There are three columns: concurrency 1, 2 and 4.

Commands to run:

$ ../pillow-perf/testsuite/run.py scale crop allocate convert -n 21

$ (../pillow-perf/testsuite/run.py scale crop allocate convert -n 21
   & ../pillow-perf/testsuite/run.py scale crop allocate convert -n 21)

$ (../pillow-perf/testsuite/run.py scale crop allocate convert -n 21
   & ../pillow-perf/testsuite/run.py scale crop allocate convert -n 21
   & ../pillow-perf/testsuite/run.py scale crop allocate convert -n 21
   & ../pillow-perf/testsuite/run.py scale crop allocate convert -n 21)

Each odd line (darker background) is the old code, each even line (white background) is the new. In each test the new code is faster, up to 2 times. Also, for operations which are not limited to memory bandwidth (resizing), performance degradation is most noticeable for the old code.

For example:
before: to 2048x1280 lzs on concurrency 4 degrades in 1.45 times
after: to 2048x1280 lzs on concurrency 4 degrades in 1.13 times
before: to 5478x3424 bil on concurrency 4 degrades in 1.56 times
after: to 5478x3424 bil on concurrency 4 degrades in 1.22 times

	conc. 1	conc. 2	conc. 4
Scale 2560×1600 RGB image
to 2048x1280 bil before	340.68 Mpx/s	270.17 Mpx/s	155.88 Mpx/s
after	420.19 Mpx/s	387.11 Mpx/s	254.74 Mpx/s
to 2048x1280 bic	273.29 Mpx/s	238.47 Mpx/s	155.75 Mpx/s
	325.72 Mpx/s	320.75 Mpx/s	245.83 Mpx/s
to 2048x1280 lzs	212.49 Mpx/s	202.35 Mpx/s	146.63 Mpx/s
	245.11 Mpx/s	241.67 Mpx/s	216.28 Mpx/s
to 5478x3424 bil	70.36 Mpx/s	62.94 Mpx/s	45.03 Mpx/s
	76.03 Mpx/s	73.74 Mpx/s	62.37 Mpx/s
to 5478x3424 bic	60.20 Mpx/s	57.15 Mpx/s	43.46 Mpx/s
	64.48 Mpx/s	63.07 Mpx/s	55.41 Mpx/s
to 5478x3424 lzs	48.56 Mpx/s	46.68 Mpx/s	40.22 Mpx/s
	51.02 Mpx/s	50.05 Mpx/s	45.36 Mpx/s
Crop 2560×1600 RGB image
2304x1440 before	1335.08 Mpx/s	735.25 Mpx/s	427.34 Mpx/s
after	2405.13 Mpx/s	1399.92 Mpx/s	863.79 Mpx/s
2816x1760	821.65 Mpx/s	439.39 Mpx/s	213.60 Mpx/s
	1178.72 Mpx/s	676.13 Mpx/s	334.77 Mpx/s
Allocate 2560×1600 RGB image
mode L before	11317.44 Mpx/s	4095.32 Mpx/s	1550.39 Mpx/s
after	22025.47 Mpx/s	6941.36 Mpx/s	2882.53 Mpx/s
mode LA	1713.02 Mpx/s	864.66 Mpx/s	413.73 Mpx/s
	5541.89 Mpx/s	2630.51 Mpx/s	1268.92 Mpx/s
mode RGB	1722.47 Mpx/s	893.53 Mpx/s	408.22 Mpx/s
	5558.03 Mpx/s	3061.27 Mpx/s	1326.43 Mpx/s
mode RGBA	1715.93 Mpx/s	892.97 Mpx/s	408.99 Mpx/s
	5572.45 Mpx/s	3049.86 Mpx/s	1332.39 Mpx/s
Convert 2560×1600 RGB image
RGB to L before	658.94 Mpx/s	638.49 Mpx/s	420.15 Mpx/s
after	692.74 Mpx/s	680.85 Mpx/s	629.48 Mpx/s
RGBA to LA	464.45 Mpx/s	412.19 Mpx/s	250.98 Mpx/s
	568.89 Mpx/s	562.02 Mpx/s	404.35 Mpx/s
RGBa to RGBA	869.12 Mpx/s	479.30 Mpx/s	253.37 Mpx/s
	1451.37 Mpx/s	814.17 Mpx/s	422.22 Mpx/s
RGBA to RGBa	916.94 Mpx/s	486.46 Mpx/s	252.86 Mpx/s
	1483.58 Mpx/s	807.89 Mpx/s	413.37 Mpx/s

[homm]

Changed base to master after merging #2654