CCCL C: Avoid passing structs by value in C APIs

[leofang]

Currently the C APIs take structs like cccl_iterator_t and cccl_op_t by value. It would be better for both performance and codegen reasons to take pointers to these structs instead. The change will also need to propagate to Python (cuda.parallel).

[bernhardmgruber]

Please verify the claim with a benchmark. Pass by value for small structs has become better over the decades, since the parameter cannot alias and may be passed through registers. It allows more optimizations.

[leofang]

Perf for now is only a potential concern that someone can help verify later (or maybe @tpn already knows! 🤩). Mainly we are using the C library from Python, and my understanding is it involves two copies of the same struct:

1. Preparing the struct data in Python
2. Preparing an FFI call (this copy is alloca so on the stack)

So if we want to do this microbenchmarking we need to do this through Python/FFI, not just writing a C code.

But perf was not the main issue I want to address 😅 We maintain an internal Python binding generator that eventually can be used to remove the ctypes hack in cuda.parallel (#3854). But it is a lot cleaner as far as codegen is concerned to pass arguments by pointer instead of by value. Saving an extra copy is only a by-product.