Use Capsule for WEBP saving

[homm]

Adoption of #8341 ideas

Improvements

• Doesn't create extra image copy (.tobytes("raw", rawmode))
• Uses the same more appropriate API .has_transparency_data for save and save_all.
• Supports saving RGBX mode without conversion
• Adds ImagingSectionEnter/Leave for WebPAnimEncoder

[Yay295]

You're essentially using _VALID_WEBP_MODES in two locations, but you removed that constant. I assume you inlined them when _VALID_WEBP_MODES and _VALID_WEBP_LEGACY_MODES were each only used in one place, but it seems like _VALID_WEBP_MODES could be kept. Also it could just be a Set instead of a Dict with True values.

[homm]

@Yay295 I've removed this constants primary since it's mistyping. Now I'm added separate _convert_frame routine for this.

[radarhere]

I expected these changes to improve speed, but I'm not finding a definitive difference when I test. Does that mirror your experience?

Are the improvements here more about theory and tidying up?

[Yay295]

I don't know how much of an effect it has, but there is some overhead when using Capsules because they use strings as identifiers. So there's some added string comparisons behind the scenes.

[homm]

@radarhere

Are the improvements here more about theory and tidying up?

All improvements are listed in the PR description. While elimination of .tobytes() doesn't affect execution time much, it helps to avid extra image copy, which reduces memory footprint. See "Maximum resident set size" below. Also, ImagingSectionEnter/Leave in WebPAnimEncoder unlocks GIL in the current thread, which improves IO in other threads.

from io import BytesIO
from PIL import Image
for i in range(4):
    with BytesIO() as f:
        Image.new('RGB', (8192, 8192)).save(f, 'webp', method=0, quality=1)

# AMD Cpu

## Main branch
$ /usr/bin/time -v python test.py
  Command being timed: "python test.py"
  User time (seconds): 19.39
  System time (seconds): 1.16
  Percent of CPU this job got: 99%
  Elapsed (wall clock) time (h:mm:ss or m:ss): 0:20.63
  Maximum resident set size (kbytes): 1296972
  Major (requiring I/O) page faults: 18
  Minor (reclaiming a frame) page faults: 1484790
  Voluntary context switches: 19
  Involuntary context switches: 719

## webp-capsule branch
$ /usr/bin/time -v python test.py 
  Command being timed: "python test.py"
  User time (seconds): 19.46
  System time (seconds): 0.78
  Percent of CPU this job got: 99%
  Elapsed (wall clock) time (h:mm:ss or m:ss): 0:20.27
  Maximum resident set size (kbytes): 1100660
  Major (requiring I/O) page faults: 0
  Minor (reclaiming a frame) page faults: 1090899
  Voluntary context switches: 1
  Involuntary context switches: 1180


# Local Docker (M1 Pro)

## Main branch
$ /usr/bin/time -v python test.py
    Command being timed: "python test.py"
    User time (seconds): 4.51
    System time (seconds): 0.35
    Percent of CPU this job got: 99%
    Elapsed (wall clock) time (h:mm:ss or m:ss): 0:04.89
    Maximum resident set size (kbytes): 851160
    Major (requiring I/O) page faults: 9
    Minor (reclaiming a frame) page faults: 674262
    Voluntary context switches: 495
    Involuntary context switches: 88

## webp-capsule branch
$ /usr/bin/time -v python test.py
    Command being timed: "python test.py"
    User time (seconds): 4.18
    System time (seconds): 0.24
    Percent of CPU this job got: 99%
    Elapsed (wall clock) time (h:mm:ss or m:ss): 0:04.44
    Maximum resident set size (kbytes): 648432
    Major (requiring I/O) page faults: 9
    Minor (reclaiming a frame) page faults: 286806
    Voluntary context switches: 390
    Involuntary context switches: 10

[homm]

@Yay295

some overhead when using Capsules because they use strings as identifiers

It's literally nothing comparing to the copying megabytes of image data to the byte string. But as you can see above, even the copying doesn't affect encoding time much.