Avoid redundant ptx generation for maximum specififed compute capability

[trevor-m]

Use

-gencode=arch=compute_xy,code=\"sm_xy,compute_xy\"

instead of

-gencode=arch=compute_xy,code=compute_xy
-gencode=arch=compute_xy,code=sm_xy

for the highest arch, which avoids generating ptx a second time. This slightly improves nvcc compilation time.

[github-actions]

Unit Test Results

     763 files  +  28       763 suites  +28   8h 34m 39s ⏱️ + 12m 40s
     768 tests ±    0       678 ✔️ ±    0       90 💤 ±    0  0 ❌ ±0 
17 698 runs  +738  12 663 ✔️ +468  5 035 💤 +270  0 ❌ ±0 

Results for commit 4b4b4bf. ± Comparison against base commit 98db066.

[github-actions]

Unit Test Results (with flaky tests)

     867 files  +  46       867 suites  +46   9h 5m 29s ⏱️ + 11m 38s
     768 tests ±    0       678 ✔️ ±    0       90 💤 ±    0  0 ❌ ±0 
20 226 runs  +980  14 119 ✔️ +512  6 107 💤 +468  0 ❌ ±0 

Results for commit 4b4b4bf. ± Comparison against base commit 98db066.

[maxhgerlach]

LGTM, thanks for the contribution @trevor-m!

Test failures for framework head versions are unrelated and already fixed on master.

nvcc docs show a slightly different syntax code=[sm_xy,compute_xy], but code=\"sm_xy,compute_xy\" works just the same. https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#extended-notation