GPU-based FFT + CMake cleanup + linalg::matrix_product

[mhaseeb123]

What's new:

• Adding optimized single-source FFT code for CPUs and GPUs
• Cleanup CMake files - Tested with nvc++/23.7 and nvc++/23.1 with both -stdpar=multicore and -stdpar=gpu
• Removed the --gcc-toolchain flag from the CMake files to encourage the use of localrc file (already setup for nvc++/23.7 btw or can be done by exporting GCCLOCALRC=/path/to/localrc) instead of the bug-prone flag.
• Added use of std::experimental::linalg library for FFT validation.
• Performance analysis for the FFT codes remains until PM GPUs free up for use. Hoping for the best though 🤞🏼
• FIXME: clang-format needed on all source and CMake files.

Note: FFT apps give a linker error for libcublas with nvc++/23.1. Added a commented out line in its CMake file which can be uncommented to add the libcublas path if needed. Interestingly nvc++/23.7 does not complain about libcublas.

[weilewei]

The code looks good overall to me! Thanks for the contribution!

One major concern is about extensive usage of macros. As our codebase grow, we probably need a better documentation for how does each macro is used. In a more ideal world, less macros, the better.