Backport: add and use preprocessors macros for loop unrolling (#32499)

[fwyzard]

Define preprocessor macros that expand to pragmas controlling loop unrolling. Two set of macros are defined:

• CMS_UNROLL_LOOP and friends are applied on both host and device compilation
• CMS_DEVICE_UNROLL_LOOP and friends are only applied in device compilation, and expand to nothing on the host

The supported compilers are

• CUDA nvcc
• HIP (untested)
• gcc
• clang

Use CMS_UNROLL_LOOP instead of #pragma unroll.