Reintroduce support for GPUs with Compute Capability 5.0

[giuliopaci]

The changes reintroduce support for Compute Capability 5.0, as described in #1765.

[thc1006]

This patch still applies cleanly to v4.8.0, and I can confirm it works on real hardware. I have a Quadro K2200 (Maxwell, cc 5.0) and hit the "no kernel image" error trying to run faster-whisper, which is how I ended up here.

I built v4.8.0 with the patch and CUDA_ARCH_LIST="5.0" on CUDA 12.9 / cuDNN 9.10. It compiles without complaint, the sm_50 kernels show up in libctranslate2.so (checked with cuobjdump), and faster-whisper transcribes correctly on the card. The AWQ __CUDA_ARCH__ < 530 guard is doing its job. For anyone wondering whether the GPU is even worth it on these old cards, float32 on the K2200 ran about 4-5x faster than CPU int8 on longer audio (int8 and fp16 aren't available on sm_50, so float32 is the only CUDA option).

The one practical reason to consider merging sooner rather than later is the toolchain window. CUDA 12.x is the last line whose nvcc still builds sm_50, and cuDNN 9.11 already drops Maxwell, so CUDA 12 is realistically the last chance for this to land in a supported way.

@giuliopaci, thank you for the fix. I put together a small repo with a reproducible build and the benchmark numbers (https://github.com/thc1006/ct2-maxwell-final) for others with these cards, though the actual code there is yours. If a CI job that publishes an sm_50 wheel would help make this easier to take on, I'd be glad to contribute one.