Refactor kernels to use dense device view

[upsj]

Using https://github.com/upsj/ginkgo-refactor to refactor headers and source files

Still WIP, since some of the replacements were incorrect

Running this requires compiling LLVM with PR llvm/llvm-project#177442
and compiling ginkgo-refactor based on it and then running

# for core source files
clang-tidy --load=./build/libRefactorDenseToView.so --checks=-*,gko-refactor-core-dense-to-view --fix file.cpp
# for kernel source files *_kernels.cpp
clang-tidy --load=./build/libRefactorDenseToView.so --checks=-*,gko-refactor-kernels-dense-to-view --fix file.cpp
# for kernel header files *_kernels.hpp
clang-tidy --load=./build/libRefactorDenseToView.so --checks=-*,gko-refactor-kernel-decls-dense-to-view --fix file.hpp

[upsj]

Issues found so far:

• some header files get mangled (these are few enough we can fix by hand, are those race conditions in parallel replacements? should be ignored by clang-tidy)
• incorrect application of replacements to some kernel_launch headers
• Dense::get_stride() replacement failing
• need to add/replace kernel_launch mappings for device views
• "missing" omp.h or hip_runtime.h headers preventing clang-tidy from running
• dpcpp support (attempt compiling against icpx LLVM?)

[upsj]

The get_stride replacement was simply missing. Creating the compile_commands.json with clang as the host compiler should fix the missing header issues. I'll try to compile against dpcpp later. Avoiding replacements in header files and member functions should avoid the kernel_launch issues, also we could skip functions that don't have the executor as first parameter (those are kernels)

[MarcelKoch]

Maybe it would make sense to add the clang formatting plugin to this PR? That could make it easier to talk about issues with the plugin. We could later decide if the plugin is also distributed in ginkgo (e.g. in the scripts dir) or not.

[upsj]

Not sure I follow, the clang-format step is just a call of clang-format afterwards. Since this is a one-time refactoring, I would keep it separate for now

[upsj]

The automated refactoring needed some significant fixes (I'll use this as a running list until I push it)

• Some macros used names like _type different from the template parameters they represented, which caused incorrect replacements. I fixed them manually
• The restriction to functions that have an executor as the first argument was too strict, it excluded functions like solve_qy in cb_gmres_kernels.cpp
• I needed to replace matrix::Dense by matrix::view::dense also in kernel_launch mapping functions, but I left the accessor in place, since it is a smaller type than the view, as it doesn't store the size. We can discuss this separately
• Some replacements in macros didn't work properly and needed to be fixed manually
• Since template deduction can deal correctly with the implicit conversion from Dense* to const Dense*, but the same isn't true for dense views, I needed to add a as_const function which casts explicitly.
• The same is true at the call site, when the core code passes in a mutable pointer, but the kernel takes a const argument
• Some pointers inside kernels could also be nullptr, I replaced them by std::optional