Try switching from rustacuda to cust

[juntyr]

rustacuda is no longer maintained, and includes all of rustacuda and most of the custom patches we are using in rust-cuda.

However, the following incompatibilities remain:

• cust_core::DeviceCopy now has Copy as a supertrait. This is both too weak and too strong for rust-cuda. There is no workaround. If cust doesn't change this requirement, no migration can occur. DeviceCopy should not require Copy Rust-GPU/Rust-CUDA#124
• Stream::wait_event takes an owned Event and Events cannot be cloned or inspected. This could be fixed by falling back to raw FFI for all usages. wait_event method in Stream Rust-GPU/Rust-CUDA#110
• Stream::add_callback no longer provides access to the current stream status. The callback might just not be scheduled anymore if an error occurred (which could result in a never-resolving async future), or do something even worse. This could be fixed by falling back to raw FFI. https://stackoverflow.com/questions/56448390/how-to-recover-from-cuda-errors-when-using-cudalaunchhostfunc-instead-of-cudastr