SYCL optimization for cache friendly data structure

[do-jason]

• Merge mdlReal and mdlImag to a single array with doubled size for cache efficient access
• Make required change and some cleanups for this mdlReal/mdlImag to mdlComplex in the required files in src/acc/ directory
• Change no_tex2D and no_tex3D to this mdlReal and mdlImag merging optimization (src/acc/sycl/sycl_kernels/sycl_utils.h)
• Remove unused complex2D and complex3D
• Change the location of group_barrier to remove unnecessary expensive barrier instruction

Compiling and running are checked for legacy, CPU, and CUDA acceleration code path.

[biochem-fan]

Thank you very much for your continued contribution.

The change looks OK to me but I do not have time to test this on many datasets.
I will tentatively merge this to the ver5.0 branch and see if anyone reports bugs.
If nothing comes in, I will merge this to master and include in 5.0.1.