[WIP] Add remove-able warnings for Beta modules or functions

[NicolasHug]

Some POC for pytorch/pytorch#72317

#######################################################
# For entire module like torchvision.models.detection #
#######################################################

# No warning - anything before torchvision.models.detection
import torchvision
import torchvision.models
from torchvision import models
from torchvision.models import resnet50
from torchvision.models import segmentation

# Any of these will warn:
from torchvision.models import detection
from torchvision.models.detection import ssd
from torchvision.models.detection.ssd import ssd300_vgg16

torchvision.models.enable_beta_detection()  # Now all warnings are disabled
from torchvision.models import detection
from torchvision.models.detection import ssd
from torchvision.models.detection.ssd import ssd300_vgg16


####################
# For single class #
####################

from contextlib import suppress
from torchvision.io import VideoReader
with suppress(RuntimeError):  # VideoReader raises a RuntimeError, but that's unrelated
    VideoReader(path="path")  # raises warning

with suppress(RuntimeError):
    torchvision.io.enable_beta_video_api()
    VideoReader(path="path")  # no warning

[NicolasHug]

Changes here are irrelevant - this is specific to pass our tests

[NicolasHug]

Here we mark the whole class as Beta. A warning is raised unless torchvision.io.enable_beta_video_api() is called first.

Functions could be treated similarly. We could have a more general decorator if needed.
This can also be used to mark single attributes or single parameters as Beta.

[NicolasHug]

This adds the detection attribute lazily. If we imported it like the others (e.g. quantization above), users would get the "beta detection warning" even when just importing torchvision.models. See more details about __getattr__ at https://peps.python.org/pep-0562/

Note: the detection attribute is needed when doing e.g.

import torchvision.models as M
M.detection  # <- directly accessing the attribute

but in a lot of patterns this gets bypassed, typically when doing

from torchvision.models.detection import ssd

or even just

import torchvision.models.detection

[NicolasHug]

Here we mark the entire torchvision.models.detection as Beta and accessing anything within its namespace will throw a warning unless torchvision.models.enable_beta_detection() is called.