Add CUDA runtime compilation support using nvrtc

[9prady9]

• Documetation for new functions (src/backend/cuda/nvrtc/cache.hpp)
• Add checks for compute versions not supported in CUDA 10

Moved the following functions to use runtime compilation while refining the API defined in cache.hpp

• convolve
• scan
• scan_by_key
• separable convolution
• transpose
• where

All required headers(listed below) for runtime compilation will be embedded into the built library. Therefore, developers writing kernels just have to #include any required files as usually inside the cuh kernel file. Check transpose function to get an idea of how it is being done.

• backend.hpp
• complex.hpp
• jit.cuh
• math.hpp
• ops.hpp
• optypes.hpp
• Param.hpp
• shared.hpp
• types.hpp

The eventual goal is to use math.hpp even inside jit kernels and remove
the code-path controlled by isJIT parameter of buildKernel function. But this has been deferred for another PR.

Notes:

1. Just this change brought down afcuda.so file size by 200MB for single compute version 61. Hopefully, we will see drastic reduction in our final binary once all feasible functions are ported to runtime compilation.
2. CUB can't be included into this framework since it includes some system headers, there is an open issue regarding this on the corresponding repository.
3. Functions using thrust can only be ported if thrust calls and raw kernels are cleanly separated. Thrust API involving CUDA runtime constructs are basically the blockers.

[9prady9]

Some other CUDA tests failed, may be which use JIT. I am looking into those failures.

[pavanky]

This is nice. good to use this for full fledged half precision support.

[9prady9]

@arrayfire/core-devel I am still debugging some failures on the linux-cuda ci job, but I think it is ready for reviews.

[umar456]

Minor comments. I have suggested an API change which may be more manageable and avoids creating a map on load.

Addressed all feedback

[9prady9]

thats an odd error on linux, that didn't happen prior to rebase I did today. trying a fresh build on my machine.

Update: was able to reproduce this on fresh build.

[9prady9]

I have noticed __int64 type failures on windows, will debug them soon.

[umar456]

A couple of small things here and there. Great comments. We need to strive to document more of our internal functions.

[umar456]

We need to do this but I am worried that there are going to be a few combinations of template parameters that are valid but will not be tested. We need be vigilant to test all types are parameters.

[9prady9]

I have rebased/squashed all CUDA work. Will push once OpenCL changes are ready.

[umar456]

You should run Bloaty McBloatface before and after on this if its not too difficult. It would be interesting to see.

[9prady9]

bloaty output

nvrtc ./src/backend/cuda/libafcuda.so.3.7.0
     VM SIZE                         FILE SIZE
 --------------                   --------------
 100.0%   119Mi TOTAL               529Mi 100.0%
  

master ./src/backend/cuda/libafcuda.so
   VM SIZE             FILE SIZE
 --------------          --------------
 100.0%  180Mi TOTAL        640Mi 100.0%

Addressed feedback