[Feature] Enable non-timm encoders with explicit embed_dim

[kitewatermelon]

Description

Fixes two failure modes that prevented custom (non-timm) backbones from being used with SSL methods:

1. Missing .embed_dim attribute
Many encoders (e.g. MONAI ViT) do not expose a .embed_dim attribute that timm models provide automatically. Passing such a module previously raised an AttributeError at construction time.

2. Hardcoded 2D dummy forward pass
Several CLS-token methods inferred embed_dim by running backbone(torch.zeros(1, 3, 224, 224)). This silently breaks any 3D encoder (e.g. volumetric MRI with shape (1, 1, 96, 96, 96)).

Fix — explicit embed_dim parameter across 13 CLS-token methods:

# timm string → embed_dim inferred automatically (no change for existing users)
model = SimCLR("vit_base_patch16_224", projector_dims=(2048, 256))

# Custom nn.Module → embed_dim provided explicitly
model = SimCLR(my_monai_vit, projector_dims=(2048, 256), embed_dim=768)
model = SimCLR(my_3d_vit,    projector_dims=(2048, 256), embed_dim=512)

Also — encoder_name API unification:
MaskedEncoder, IJEPA, MAE, and LeJEPA previously accepted the encoder argument as model_or_model_name. Renamed to encoder_name to match all other methods in the library.

Affected methods: BarlowTwins, BYOL, DINO, MoCov2, MoCov3, NNCLR, PIRL, SimCLR, SimSiam, SwAV, TiCO, VICReg, WMSE, MaskedEncoder, IJEPA, MAE, LeJEPA

Note: Although model_or_model_name may be a more descriptive name, encoder_name was chosen to match the existing convention already used across the majority of the codebase.

Together, encoder_name and embed_dim make the methods API more predictable and consistent for users bringing custom backbones.

Checklist

• I have read the Contributing document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR to the RELEASES.rst file.