fix(mask2former): register input_projections as nn.ModuleList (fixes #49698) (#49703)
fix(mask2former): register input_projections as nn.ModuleList
Mask2FormerTransformerModule kept input_projections in a plain Python
list, so the input projection layers were never registered as PyTorch
submodules: their parameters were excluded from model.parameters(),
optimizers and the state dict, and model.to(device) left them on CPU
causing 'input type ... and weight type ... should be the same'
RuntimeErrors on CUDA. The pixel decoder already uses nn.ModuleList for
the same structure (line 1266-1276) and the original Mask2Former
implementation registers them the same way.
Fixes #49698