ROCm build fails with torch >=2.11: c10::hip::HIPCachingAllocator has not been declared

[LunNova]

PyTorch's MasqueradingAsCUDA cleanup dropped the c10::hip namespace, so the conditional namespace selection for cache clearing no longer works. I think everything can just use the cuda namespace.

PyTorch PR:
pytorch/pytorch#174087

Code that needs changed:
https://github.com/pierotofy/OpenSplat/blob/1f4dc9b29752341220b00a87d137a4f4b81b94e7/model.cpp#L15-L19
https://github.com/pierotofy/OpenSplat/blob/1f4dc9b29752341220b00a87d137a4f4b81b94e7/model.cpp#L487-L491