Decide on Nvidia GPU support for 2025.06

We should have a discussion on which Nvidia Compute capabilities we support in 2025.06. This should take into account that the traditional forwards-compatiblity (which allowed os to build for e.g. 8.0 and have all 8.X cards supported) that existed with major versions is now no longer a guarantee.

There are family-specific features, and architecture-specific features. E.g. compiling for compute_90a enables features specific to the 9.0 architecture, that are not guaranteed to run on any newer 9.X architecture. Similarly, there are family-sepcific features that can be enabled when compiling for e.g. compute_100f, which are compatible only with 10.0 and 10.3 cards.

This is the relevant part of the docs that explain these architecture and family-specific features.

One option could be that we don't compile with support for family or architecture specific features. However, right now, there is only one 9.X architecture, namely 9.0, so it feels a bit silly not to use the full 9.0 capability. Furthermore, we've already seen issues in practice with this: many libs in cuDNN only support 9.0a, and not the 'plain' 9.0. Thus, if we ship this cuDNN, we cannot claim full compatiblity with 9.0 (i.e. any 9.X card), since we are 'only' compatible with 9.0a (i.e. any 9.0 card).

Note that, in theory, this discussion also affects 2023.06 since we claim to support '9.0' there. However, the cuDNN versions there did not show failures in the CUDA sanity checks, so I assume that all of those still shipped actual 9.0 code for all their libs. I.e. at least for now, that stack should thus be fully 9.0-compatible.