Get rid of unneeded version checks

[kshyatt]

According to the release notes, CUDA 11.4 ships with CUBLAS 11.6.5, so these checks for version() < 11 aren't reachable.

[github-actions]

CUDA.jl Benchmarks

Benchmark suite	Current: ca96e09	Previous: 5645ddc	Ratio
latency/precompile	42854033083 ns	48881022269.5 ns	0.88
latency/ttfp	7148038767 ns	7292511536 ns	0.98
latency/import	3387489703 ns	3494844990 ns	0.97
integration/volumerhs	9624264.5 ns	9618776.5 ns	1.00
integration/byval/slices=1	147256 ns	147140 ns	1.00
integration/byval/slices=3	425877 ns	425273 ns	1.00
integration/byval/reference	145298 ns	145049 ns	1.00
integration/byval/slices=2	286513 ns	286349 ns	1.00
integration/cudadevrt	103694 ns	103459 ns	1.00
kernel/indexing	14436 ns	14269 ns	1.01
kernel/indexing_checked	14861 ns	14648 ns	1.01
kernel/occupancy	777.5090909090909 ns	720.354609929078 ns	1.08
kernel/launch	2285.5555555555557 ns	2197.3333333333335 ns	1.04
kernel/rand	18461 ns	14675 ns	1.26
array/reverse/1d	19857 ns	19825.5 ns	1.00
array/reverse/2d	24929 ns	25322 ns	0.98
array/reverse/1d_inplace	10796 ns	10664 ns	1.01
array/reverse/2d_inplace	12244 ns	12288 ns	1.00
array/copy	21381 ns	21417 ns	1.00
array/iteration/findall/int	158440 ns	160202.5 ns	0.99
array/iteration/findall/bool	139170 ns	140858 ns	0.99
array/iteration/findfirst/int	153951 ns	153807 ns	1.00
array/iteration/findfirst/bool	154987 ns	154741 ns	1.00
array/iteration/scalar	72845 ns	72443 ns	1.01
array/iteration/logical	216325 ns	219718.5 ns	0.98
array/iteration/findmin/1d	41346.5 ns	41952 ns	0.99
array/iteration/findmin/2d	94176 ns	94212 ns	1.00
array/reductions/reduce/1d	35853 ns	36691.5 ns	0.98
array/reductions/reduce/2d	41462 ns	49814 ns	0.83
array/reductions/mapreduce/1d	34029 ns	34861 ns	0.98
array/reductions/mapreduce/2d	41368 ns	41414 ns	1.00
array/broadcast	21097 ns	20872 ns	1.01
array/copyto!/gpu_to_gpu	12972 ns	13686 ns	0.95
array/copyto!/cpu_to_gpu	214298 ns	211802 ns	1.01
array/copyto!/gpu_to_cpu	283270 ns	245699 ns	1.15
array/accumulate/1d	109578 ns	109690 ns	1.00
array/accumulate/2d	80665 ns	80775 ns	1.00
array/construct	1281.8 ns	1289.6 ns	0.99
array/random/randn/Float32	43925 ns	44847 ns	0.98
array/random/randn!/Float32	25118 ns	26739.5 ns	0.94
array/random/rand!/Int64	27236 ns	27115 ns	1.00
array/random/rand!/Float32	8871 ns	8873 ns	1.00
array/random/rand/Int64	30222 ns	38396 ns	0.79
array/random/rand/Float32	13438 ns	13462 ns	1.00
array/permutedims/4d	61819 ns	61480 ns	1.01
array/permutedims/2d	55360 ns	55970.5 ns	0.99
array/permutedims/3d	56269 ns	56435 ns	1.00
array/sorting/1d	2766536 ns	2776551 ns	1.00
array/sorting/by	3354483 ns	3368660 ns	1.00
array/sorting/2d	1082927.5 ns	1086044 ns	1.00
cuda/synchronization/stream/auto	1025.9 ns	1062.2 ns	0.97
cuda/synchronization/stream/nonblocking	7550.6 ns	6529.8 ns	1.16
cuda/synchronization/stream/blocking	798.6632653061224 ns	804.8736842105263 ns	0.99
cuda/synchronization/context/auto	1182.1 ns	1171.2 ns	1.01
cuda/synchronization/context/nonblocking	8394.6 ns	6701.6 ns	1.25
cuda/synchronization/context/blocking	904.4 ns	921.8888888888889 ns	0.98

This comment was automatically generated by workflow using github-action-benchmark.

[maleadt]

Hm, looks like the dist-upgrade of gpuci broke some stuff, and c1009e6 isn't sufficient as a workaround. It will have to wait until next week before I have the time to investigate though...