OnlineKNN: OnlineQueue buffers live on CPU, causing device mismatches and cpu->gpu transfrs

[sami-bg]

OrderedQueue buffers are always created on CPU because OnlineQueue.setup() never moves them to the model's device after creation. This causes problems in a few places:

NNCLR is broken, because _find_nearest_neighbors in forward.py does torch.mm(query_norm, support_norm.t()) where the query is on cuda but the support set from the queue is on cpu.

Swav and KNN have band-aids that cover this: swav_forward does queue.get().clone().detach().to(proj1.device), and OnlineKNN._compute_knn_predictions checks if cached_features.device != features.device and moves. These work but they're copying the entire queue from CPU to GPU every time the data is consumed. for NNCLR's support set that's 65K x 256 (~64MB) every training step, and for KNN it's 20K x 512 every validation batch.

The fix is OnlineQueue.setup(), after creating or resizing the OrderedQueue, move it to pl_module's device:

device = next(pl_module.parameters()).device
self._shared_queues[self.key].to(device)