Allow return statements for GPU-only kernels

[pxl-th]

It's a bit annoying to not be able to use return statements.
Currently you have to do something like this, but it looks weird:

@kernel function ker!(args...)
    i = @index(Global)
    _ker!(i, args...)
end

function _ker!(i, args...)
    ...
end

[github-actions]

Benchmark Results

	main	b3f6676...	main/b3f66760325ea4...
saxpy/default/Float16/1024	0.541 ± 0.0044 μs	0.53 ± 0.0041 μs	1.02
saxpy/default/Float16/1048576	0.173 ± 0.0024 ms	0.175 ± 0.0037 ms	0.993
saxpy/default/Float16/16384	3.16 ± 0.053 μs	3.13 ± 0.05 μs	1.01
saxpy/default/Float16/2048	0.708 ± 0.0062 μs	0.7 ± 0.0064 μs	1.01
saxpy/default/Float16/256	0.411 ± 0.0028 μs	0.404 ± 0.0035 μs	1.02
saxpy/default/Float16/262144	0.044 ± 0.00071 ms	0.0457 ± 0.0026 ms	0.964
saxpy/default/Float16/32768	5.87 ± 0.1 μs	5.8 ± 0.09 μs	1.01
saxpy/default/Float16/4096	1.09 ± 0.015 μs	1.09 ± 0.017 μs	1
saxpy/default/Float16/512	0.461 ± 0.0032 μs	0.451 ± 0.0037 μs	1.02
saxpy/default/Float16/64	0.382 ± 0.0027 μs	0.381 ± 0.0044 μs	1
saxpy/default/Float16/65536	11.5 ± 0.2 μs	11.5 ± 0.22 μs	1
saxpy/default/Float32/1024	0.44 ± 0.0063 μs	0.436 ± 0.006 μs	1.01
saxpy/default/Float32/1048576	0.235 ± 0.013 ms	0.231 ± 0.012 ms	1.02
saxpy/default/Float32/16384	2.66 ± 1 μs	3.17 ± 0.98 μs	0.839
saxpy/default/Float32/2048	0.551 ± 0.012 μs	0.541 ± 0.012 μs	1.02
saxpy/default/Float32/256	0.391 ± 0.0049 μs	0.39 ± 0.0065 μs	1
saxpy/default/Float32/262144	0.0581 ± 0.003 ms	0.0548 ± 0.0035 ms	1.06
saxpy/default/Float32/32768	5.33 ± 1.6 μs	6.1 ± 1.6 μs	0.874
saxpy/default/Float32/4096	0.918 ± 0.018 μs	0.91 ± 0.054 μs	1.01
saxpy/default/Float32/512	0.405 ± 0.0041 μs	0.403 ± 0.0044 μs	1
saxpy/default/Float32/64	0.381 ± 0.004 μs	0.372 ± 0.0032 μs	1.02
saxpy/default/Float32/65536	12.3 ± 1.4 μs	12.5 ± 1.7 μs	0.985
saxpy/default/Float64/1024	0.558 ± 0.013 μs	0.532 ± 0.011 μs	1.05
saxpy/default/Float64/1048576	0.485 ± 0.024 ms	0.508 ± 0.051 ms	0.955
saxpy/default/Float64/16384	6.41 ± 1.2 μs	5.15 ± 1.5 μs	1.25
saxpy/default/Float64/2048	0.922 ± 0.0082 μs	0.915 ± 0.023 μs	1.01
saxpy/default/Float64/256	0.408 ± 0.0047 μs	0.403 ± 0.0046 μs	1.01
saxpy/default/Float64/262144	0.114 ± 0.0068 ms	0.118 ± 0.008 ms	0.964
saxpy/default/Float64/32768	13.1 ± 2.2 μs	12.7 ± 1.1 μs	1.03
saxpy/default/Float64/4096	1.71 ± 0.27 μs	1.6 ± 0.24 μs	1.07
saxpy/default/Float64/512	0.443 ± 0.0054 μs	0.434 ± 0.006 μs	1.02
saxpy/default/Float64/64	0.387 ± 0.0035 μs	0.385 ± 0.003 μs	1.01
saxpy/default/Float64/65536	28.2 ± 3 μs	29.2 ± 1.5 μs	0.965
saxpy/static workgroup=(1024,)/Float16/1024	1.91 ± 0.025 μs	1.91 ± 0.022 μs	1
saxpy/static workgroup=(1024,)/Float16/1048576	0.174 ± 0.013 ms	0.161 ± 0.0071 ms	1.08
saxpy/static workgroup=(1024,)/Float16/16384	4.26 ± 0.14 μs	4.17 ± 0.13 μs	1.02
saxpy/static workgroup=(1024,)/Float16/2048	2.09 ± 0.032 μs	2.09 ± 0.038 μs	0.998
saxpy/static workgroup=(1024,)/Float16/256	2.54 ± 0.028 μs	2.55 ± 0.024 μs	0.996
saxpy/static workgroup=(1024,)/Float16/262144	0.0422 ± 0.0019 ms	0.0438 ± 0.003 ms	0.965
saxpy/static workgroup=(1024,)/Float16/32768	6.73 ± 0.26 μs	6.58 ± 0.25 μs	1.02
saxpy/static workgroup=(1024,)/Float16/4096	2.39 ± 0.032 μs	2.4 ± 0.029 μs	0.999
saxpy/static workgroup=(1024,)/Float16/512	2.98 ± 0.052 μs	3 ± 0.048 μs	0.992
saxpy/static workgroup=(1024,)/Float16/64	2.24 ± 0.041 μs	2.26 ± 0.039 μs	0.991
saxpy/static workgroup=(1024,)/Float16/65536	12.7 ± 0.53 μs	12.6 ± 0.53 μs	1.01
saxpy/static workgroup=(1024,)/Float32/1024	1.93 ± 0.021 μs	1.95 ± 0.031 μs	0.993
saxpy/static workgroup=(1024,)/Float32/1048576	0.236 ± 0.013 ms	0.254 ± 0.022 ms	0.93
saxpy/static workgroup=(1024,)/Float32/16384	4.18 ± 0.72 μs	4.22 ± 0.72 μs	0.991
saxpy/static workgroup=(1024,)/Float32/2048	2.08 ± 0.03 μs	2.08 ± 0.031 μs	1
saxpy/static workgroup=(1024,)/Float32/256	2.4 ± 0.037 μs	2.44 ± 0.041 μs	0.984
saxpy/static workgroup=(1024,)/Float32/262144	0.062 ± 0.0032 ms	0.0613 ± 0.0017 ms	1.01
saxpy/static workgroup=(1024,)/Float32/32768	7.62 ± 1.1 μs	7.33 ± 0.6 μs	1.04
saxpy/static workgroup=(1024,)/Float32/4096	2.34 ± 0.043 μs	2.37 ± 0.047 μs	0.989
saxpy/static workgroup=(1024,)/Float32/512	2.4 ± 0.04 μs	2.45 ± 0.039 μs	0.98
saxpy/static workgroup=(1024,)/Float32/64	2.63 ± 8.1 μs	2.65 ± 7.7 μs	0.992
saxpy/static workgroup=(1024,)/Float32/65536	16.5 ± 1.9 μs	15.4 ± 1.1 μs	1.07
saxpy/static workgroup=(1024,)/Float64/1024	2.03 ± 0.036 μs	2.03 ± 0.027 μs	1
saxpy/static workgroup=(1024,)/Float64/1048576	0.542 ± 0.054 ms	0.52 ± 0.038 ms	1.04
saxpy/static workgroup=(1024,)/Float64/16384	7.79 ± 1.2 μs	7.19 ± 0.68 μs	1.08
saxpy/static workgroup=(1024,)/Float64/2048	2.3 ± 0.047 μs	2.3 ± 0.052 μs	0.999
saxpy/static workgroup=(1024,)/Float64/256	2.4 ± 0.057 μs	2.43 ± 0.062 μs	0.986
saxpy/static workgroup=(1024,)/Float64/262144	0.122 ± 0.0092 ms	0.121 ± 0.0077 ms	1.01
saxpy/static workgroup=(1024,)/Float64/32768	15.7 ± 2 μs	15.6 ± 1.4 μs	1.01
saxpy/static workgroup=(1024,)/Float64/4096	3.04 ± 0.23 μs	2.92 ± 0.22 μs	1.04
saxpy/static workgroup=(1024,)/Float64/512	2.38 ± 0.046 μs	2.41 ± 0.052 μs	0.991
saxpy/static workgroup=(1024,)/Float64/64	2.38 ± 4.6 μs	2.4 ± 18 μs	0.991
saxpy/static workgroup=(1024,)/Float64/65536	31.1 ± 2.7 μs	0.032 ± 0.0019 ms	0.972
time_to_load	0.726 ± 0.017 s	0.75 ± 0.0098 s	0.968

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[vchuravy]

This needs tests. Making it GPU only is fine given #533

[pxl-th]

@vchuravy if that's OK, I'd also like to tag a new release after this PR is merged.