Support for 'delayed kernels'

[maleadt]

When you need to introspect the compiled kernel, e.g. to determine a launch configuration, you either have to do the whole cudaconvert and cufunction dance manually, or use the hacky config=callback argument to @cuda. Both are pretty cumbersome, so here I introduce an alternative: @cuda delayed=true kernel(args...), returning a callable object you can then just introspect and finally call using kernel_object(args...; threads=..., blocks=..., shmem=...).

cc @vchuravy, as KA probably uses the lower-level interface.

[maleadt]

Also, I don't particularly like the delayed name, so bikeshed away!

[maleadt]

Or maybe it should be @cufunction.

[codecov]

Codecov Report

Merging #569 (d7cd82e) into master (3c9e3dd) will decrease coverage by 0.06%.
The diff coverage is 96.07%.

@@            Coverage Diff             @@
##           master     #569      +/-   ##
==========================================
- Coverage   80.34%   80.28%   -0.07%     
==========================================
  Files         116      116              
  Lines        6889     6883       -6     
==========================================
- Hits         5535     5526       -9     
- Misses       1354     1357       +3     

Impacted Files	Coverage Δ
src/compiler/execution.jl	89.92% <81.81%> (-1.62%)	⬇️
examples/pairwise.jl	78.00% <100.00%> (-0.85%)	⬇️
examples/peakflops.jl	100.00% <100.00%> (ø)
src/accumulate.jl	97.14% <100.00%> (-0.23%)	⬇️
src/gpuarrays.jl	40.90% <100.00%> (-4.93%)	⬇️
src/indexing.jl	100.00% <100.00%> (ø)
src/mapreduce.jl	97.72% <100.00%> (-0.15%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3c9e3dd...c8a8573. Read the comment docs.

[vchuravy]

KA probably uses the lower-level interface.

GPUifyLoops used to, but for KA I switched back to the simpler one. But yes this is great @cufunction is great.

[maleadt]

Changed it from delayed=true to launch=false. Every @cuda now returns the kernel object too. I decided against @cufunction because it uses values, not types.