cmake: fix CMAKE_CUDA_ARCHITECTURES default

[cebtenzzre]

Because we specify a minimum CMake version of 3.21 in the llmodel project (gpt4all-backend), we get the NEW behavior of CMP0104, which causes enable_langauge(CUDA) to set CMAKE_CUDA_ARCHITECTURES to be set based on nvcc's default.

On my machine this means CMAKE_CUDA_ARCHITECTURES gets set to 52, for my GTX 970. No idea what the v2.8.0 installers were built with, but with this change we can be sure that the next release will have the expected feature support
on 6.1+ architectures.

This is the issue that caused me to accidentally be able to reproduce the PTX incompatibility issue (because I was building for sm_52 but not sm_61), but it does not itself fix the inherent problem with the PTX code shipped with GPT4All v2.8.0 (they were built with too new of a CUDA version for many users, including all Linux users until NVIDIA 555.x comes out of Beta).

[ThiloteE]

I've done tests with Windows 10, AMD 5600 and Nvidia GTX 1060 3GB.
Nvidia driver version 555.99 (32.0.15.5599)
Cuda toolkit: cuda_12.4.1_551.78_windows

👍 2.8.0 Cuda backend does NOT crash.
👎 main with commit 41c9013: Cuda backend DOES crash.
👍 This PR at bd5dfe1: Cuda backend does NOT crash.

Model: Various. E.g.

• qwen1.5-moe-a2.7b-chat-imat-IQ4_XS.gguf
• tinyllama-1.1b-chat-v1.0.Q4_0.gguf
• phi-3 Mini Instruct Q4_0
• llama-3 Q4_0