Unify and simplify batch functionality: Multivector

[pratikvn]

This PR is the first part of the batched kernel refactoring. It removes the .inc style files, moves to .hpp, making the functions available in a common namespace. It also uses a single source for both HIP and CUDA.

TODO

• Check if there are any issues with solver includes.
• Remove existing redundant .inc files
• Differences in __launch_bounds__ between CUDA and HIP needs to be resolved

[MarcelKoch]

If you haven't noted that already, there is a naming inconsistency of the single-batch matrix-apply kernels for the reference and cuda/hip backend.
In the reference backend those are called simple|advanced_apply_kernel, while for cuda/hip they are called simple|advanced_apply.
I think that should also be unified.

[upsj]

LGTM, I'm wondering if we can get rid of some of the GKO_DEVICE_NAMESPACE:: annotations though - is there an ambiguity issue when removing them?

[pratikvn]

@MarcelKoch , yes, I plan to unify those when refactoring the matrix format kernels.

[yhmtsai]

you have used 4 for sm_oversubscription on both cuda/hip.
I assume the cuda is the correct and hip just uses it.
if you want to compute in more accurate mapping, hip should use (min_blocks_multiprocessor (4) * max_threads_per_block (256) )/32 = 32 for hip.
you will need to distinguish it by macro

[pratikvn]

I havent fully benchmarked that yet. I agree that this will have to be a macro specialized for CUDA and HIP. But will done in a future PR. It has already been noted in #1376

[pratikvn]

format!