arrayfire / arrayfire

* Add missing batch support check in sparse-dense arith ops

* Fix formatting

Using single dynamic library instead of the tuple of interface,
threading-layer & core libraries removes the linking issues in unified
backend library. This further removes issues from wrappers that use
unified backend when loading Intel MKL libraries at runtime.

With this change, we also package mkl_rt single dynamic library along
with all other required libraries.

This optimization dynamically sets the block size based on the output array
dimension. Originally we had a block size of 32x8 threads per block. This
configuration was not ideal when indexing into a long array where you
had few columns and many rows. The current approach creates blocks of
256x1, 128x2, 64x4 and 32x8 to better accommodate smaller dimensions.

This was a problem on the arrayfire-benchmark repo where the repository
is built as a subproject

Splitting the windows ci job into a separate workflow enables the
ci to re-run windows specific jobs independent of unix jobs.

Updated Ninja dependency to 1.10.2 fix release in all ci(gh-actions)

Refactored boost dependency to be installed via packages managers as
GitHub Actions is removing pre-installed versions from March 8, 2021

Update VCPKG hash to newer version to enable fast and better ports.

These are not needed since the move to getting dependencies using
fetch content module of cmake.

Refactored release source tar ball action to relfect the same as well

This results in faster re-runs of cmake command after the first run.

Also removed obsolete clblas prefix and associated variables.

unmarked CUDA_VERSION as advanced so that users may see what CUDA
toolkit is picked up

These functions are not exposed to users. They are not included when
generating installers. These functions are purely for testing certain
internal behavior given a certain combination of environment variables.

Test for unit max JIT height infinite recursion bug

When the maximum JIT tree height is one, createNaryNode goes into
infinite recursion. This effects CUDA and OpenCL backends

16fp and 64fp are optional extensions to OpenCL.  The CONFIG's only exists when the extension is available.  It is therefore better to check the availability of the extension, so that no errors are thrown (and have to treated).

+ Cleanup of compiler warnings.

#3043)

* Reduced overhead of kernel caching for OpenCL & CUDA.

The program source files memory footprint is reduced (-30%) by eliminating
comments in the generated kernel headers.  Hash calculation of each source
file  is performed at compile time and incrementally extended at runtime
with the options & tInstance vectors.  Overall performance increased up to
21%, up to the point that the GPU becomes the bottleneck, and the overhead
to launch the same (small) kernel was improved by 63%.

* Fix couple of minor cmake changes

* Move spdlog fetch to use it in bin2cpp link command

Co-authored-by: pradeep <pradeep@arrayfire.com>

- cl2.hpp header download
- clBLAS build
- clFFT build
- CLBlast build

Use clBLAS and clFFT via add_subdir instead of external project

When the above cmake option is turned ON, via the below command
```cmake
ccmake .. -DAF_BUILD_OFFLINE:BOOL=ON
```
FetchContent will look for dependencies under build tree's extern
folder and will not attempt to download any of them.

By default this option is turned OFF

- spdlog
- cub
- threads

Removed assets and test/data as submodules