Compute capability is a property of the GPU hardware and immutable for a given GPU. As you already found out, the Quadro RTX 3000 is based on the Turing architecture, with compute capability 7.5 (
Support for the Turing architecture was added in CUDA 10.0. The latest CUDA version is 11.2.2. Your GPU will work fine with any CUDA version between those endpoints. However, the various other components in your deep learning software stack may have very specific requirements as to CUDA version, so check relevant information on software requirements for those components.
I do not use PyTorch, but the “Installation” section at the start page for the project (that you linked) seems to indicate that at present, the stable 1.8.1 version of PyTorch requires either CUDA 10.2 or CUDA 11.1. I would not have the faintest clue what the trade-offs are between those two CUDA version with regard to PyTorch. From a general CUDA perspective, choosing the newer version CUDA 11.1 seems appropriate.