Support extents with arbitrary value types

[bernhardmgruber]

Currently, LLAMA uses size_t in all index calculations as is the default in the STL. We knew already from previous experience in picongpu that this increases register usage for CUDA since size_t is 64bits in CUDA and CUDA registers are 32bits in size, so each index variable uses 2 registers. Registers are a scarce resource so it would be beneficial to have control over the type used in index calculations. This PR allows this.

We saw similar development in std::mdspan (https://wg21.link/P0009) with the proposal “Make mdspan size_type controllable” (https://wg21.link/P2553). The gist was adding an additional template parameter to std::extents which gives the type used in all index calculations:

template<class SizeT, size_t... Extents>
  class extents;

For LLAMA, the main change is that llama::ArrayExtents gains an additional template parameter defining the type to use for storing extents and performing index calculations in:

using AE = llama::ArrayExtents<IndexType, 5, llama::dyn, 2>;

This change is propagated downstream into llama::ArrayExtentsDynamic<IndexType, 3>, which is a breaking change. llama::NrAndOffset is now templated on the index type as well. Furthermore ArrayExtentsStatic is renamed to ArrayExtentsNCube, since this is a more accurate description of what it does.

[codecov]

Codecov Report

Merging #488 (d6cfa5a) into develop (50bac5f) will increase coverage by 0.00%.
The diff coverage is 99.87%.

❗ Current head d6cfa5a differs from pull request most recent head da970d8. Consider uploading reports for the commit da970d8 to get more accurate results

@@           Coverage Diff            @@
##           develop     #488   +/-   ##
========================================
  Coverage    99.23%   99.24%           
========================================
  Files           63       63           
  Lines         5392     5439   +47     
========================================
+ Hits          5351     5398   +47     
  Misses          41       41           

[bernhardmgruber]

Just tested this one in AdePT and it works.