Respect strict=False when loading detection models

[datumbox]

Fixes #5835

Tested with:

from torchvision.models import detection

fns = [v for k, v in detection.__dict__.items() if callable(v) and k[0].lower() == k[0] and k[0] != "_"]
for fn in fns:
    try:
        model = fn(weights_backbone=None)
        model.load_state_dict({}, strict=False)
        print(fn.__name__, "PASS")
    except:
        print(fn.__name__, "FAIL")

Output:

fasterrcnn_resnet50_fpn PASS
fasterrcnn_resnet50_fpn_v2 PASS
fasterrcnn_mobilenet_v3_large_fpn PASS
fasterrcnn_mobilenet_v3_large_320_fpn PASS
fcos_resnet50_fpn PASS
keypointrcnn_resnet50_fpn PASS
maskrcnn_resnet50_fpn PASS
maskrcnn_resnet50_fpn_v2 PASS
retinanet_resnet50_fpn PASS
retinanet_resnet50_fpn_v2 PASS
ssd300_vgg16 PASS
ssdlite320_mobilenet_v3_large PASS

I'm not adding the test in the unit-tests to avoid slowing them down further. This gap can be resolved properly once we redesign the testing strategy for models.

[NicolasHug]

Thanks @datumbox