port cuda coo to amd coo

[yhmtsai]

This PR ports cuda coo implementation into amd coo.

HIP does not support the warp-level cooperative group.
I add some implementations of warp-level cooperative group, it works fine now but it is not same as cooperative group.
I write the similar interface to wrap the shfl, any, etc functions.
If HIP or we implement the correct cooperative group, it only need to change cooperative_group.hip.hpp and does not hurt any other implementation.
Moreover, the code is messy and the optimization is nothing (I use min(300, nnz/warp_size) for calculating the number of warp, 300 is randomly set).
The optimization is done but need to tune. AMD has 4 SIMD in each multiprocessor. I use it to calculate the number of warps.

There are two failed test (Coo.AdvancedApplyIsEquivalentToRef and Coo.AdvancedApplyToDenseMatrixIsEquivalentToRef) because they need dense hip code.
Edit: comment them until implementing hip dense kernel.
Other tests of Coo are passed.

Summary:

1. add cooperative_group for hip
2. split config and type
3. hip coo codes
4. Create common folder to handle the duplicated code of hip and cuda.
   the name.hpp.inc in common puts the subnamespace and does not include any headers.

namespace proper {
    namespace parent {
        // some constexpr variable
        #include "common/path/to/name.hpp.inc"
        // other codes
    }
}

TODO:

• Merge Initial code for the HIP executor #344
• add hip optimization

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (develop@f4ae401). Click here to learn what that means.
The diff coverage is 32.2%.

@@            Coverage Diff             @@
##             develop     #357   +/-   ##
==========================================
  Coverage           ?   64.02%           
==========================================
  Files              ?      265           
  Lines              ?    16358           
  Branches           ?        0           
==========================================
  Hits               ?    10474           
  Misses             ?     5884           
  Partials           ?        0

Impacted Files	Coverage Δ
include/ginkgo/core/base/lin_op.hpp	78.48% <ø> (ø)
core/test/utils/assertions.hpp	85.03% <ø> (ø)
include/ginkgo/core/base/types.hpp	92.59% <ø> (ø)
reference/test/matrix/hybrid_kernels.cpp	100% <ø> (ø)
include/ginkgo/core/base/exception_helpers.hpp	100% <ø> (ø)
include/ginkgo/core/base/math.hpp	90.32% <ø> (ø)
core/base/extended_float.hpp	79.12% <ø> (ø)
include/ginkgo/core/stop/criterion.hpp	84.61% <ø> (ø)
benchmark/spmv/spmv.cpp	0% <ø> (ø)
benchmark/utils/general.hpp	0% <0%> (ø)
... and 26 more

---

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 f4ae401...e0caafb. Read the comment docs.

[tcojean]

LGTM in general. I do have some comments though.

[yhmtsai]

This is updated coo vs hipsp_coo

[pratikvn]

LGTM!

[upsj]

LGTM!
Most of the changes were code moving between files and tiny refactorings anyways.
A comment and a suggestion:

1. The tests for HIP COO kernels are missing two test cases that are present in the CUDA tests
2. Could we not quickly manually implement the any, all and ballot methods for sub-warp tiles?