Batched classes todo list

[pratikvn]

This issue tracks the different aspects of the batched functionality that reviewers have suggested, but still need to be implemented:

1. Performance related:

• Handling multiple RHS, or special case for single rhs
• Workgroup, grid and block sizes for GPU kernels
• __launch_bounds__ in CUDA/HIP kernels.
• Remove unnecessary temporary stack variables to reduce register usage.
• Assign multiple threads to a row for batch::Ell SpMV. See gko::matrix::Ell kernels.
• Check unecessary synchronizations and use the necessary one (global_space, local_space etc)

2. Utility related:

• Add stride capability.

3. Code simplifications:

• Using a common launcher syntax, something similar to parfor_batch_blockwise(x, [](ibatch) { ... })
• Check if batch_item and uniform_batch can be combined into one struct.
• Check for getters and functions in classes that are unused and can be removed from the interface.
• Unify stopping criteria and loggers across device backends by moving them into common/unified
• Unify simple preconditioners into common/unified
• Simplify logger interface, and include the selection within the batch solver dispatch.
• Simplify/separate storage_config in core and kernels and use only necessary members for the kernels.
• Refactor compute_shared_storage functionality, and unify computations across the different batched solver classes.
• Clarify include files in .hpp.inc by maybe splitting the files and having a _includes.hpp.inc