Remove deprecated collate functions, model wrappers, and arguments

[gabrielfruet]

Removes the deprecated collate functions, the high-level model wrappers, SymNegCosineSimilarityLoss, and a set of deprecated arguments. All of it has warned since 1.3.0 or 1.4.0.

Breaking change — needs a minor version bump at release, which is not part of this PR.

Rebased onto master now that #1939 and #2020 have landed; the diff is this PR's own four commits.

Migration

Removed	Replacement
every collate function in lightly.data.collate (Base, Image, MultiView, SimCLR, MoCo, MultiCrop, SwaV, DINO, MAE, PIRL, MSN, SMoG, VICReg, VICRegL)	the matching lightly.transforms.*Transform plus MultiViewCollate
lightly.data.imagenet_normalize	lightly.transforms.utils.IMAGENET_NORMALIZE
from lightly.data.collate import IJEPAMaskCollator — the whole lightly.data.collate module goes, and with it the re-export #2020 left behind	from lightly.data import IJEPAMaskCollator
lightly.models.{BarlowTwins, BYOL, MoCo, NNCLR, SimCLR, SimSiam}	heads from lightly.models.modules, as in examples/
lightly.loss.SymNegCosineSimilarityLoss	lightly.loss.NegativeCosineSimilarity
MSNLoss(me_max_weight=...)	MSNLoss(regularization_weight=...)
GaussianBlur(kernel_size=..., scale=...) and the pass-throughs on the method transforms	GaussianBlur(sigmas=...)
MemoryBankModule(size=n)	MemoryBankModule(size=(n, dim))
trainer.weights_summary in the CLI config	trainer.enable_model_summary and summary_callback.max_depth
plot_augmented_images(images, collate_function)	plot_augmented_images(images, transform), taking a MultiViewTransform

Worth knowing

lightly-ssl-train and lightly-magic build a SimCLRTransform instead of an ImageCollateFunction. The augmentations are identical: every key in the collate: config namespace maps onto a SimCLRTransform argument and all of them are pinned in config.yaml. The namespace keeps its name because lightly-embed reads collate.input_size.

GaussianBlur.__init__ is keyword-only now. kernel_size was the first positional argument, so removing it would otherwise reinterpret GaussianBlur(0.3) as prob=0.3 without complaint.

Dropping the blur arguments shifts positional order on roughly 30 public transform classes, so callers passing arguments positionally past that point need keywords.

MemoryBankModule raises ValueError on a bare positive int size. Inferring the feature dimension from the first batch broke distributed training. size=0 still disables the bank, so NTXentLoss(memory_bank_size=0) is unaffected.

Testing

Full pytest suite after the rebase: 1397 passed, 226 skipped. ruff format-check and lint are clean. The one failure, tests/utils/test_dist__gather__losses.py::test_loss_dcl, is the macOS DDP hang from #1917 and fails identically on an unmodified master checkout.

mypy and the Sphinx docs build don't run on this machine — mypy 1.4.1 crashes on the numpy 2.5.1 stubs, and Sphinx wants imghdr, which Python 3.13 removed. Instead I imported every autodoc target in docs/source to confirm it still resolves.