Avoid the fancy-shmancy heap allocation for stream callbacks

[eyalroz]

At the moment, when a host (callback) function or callable is enqueued on a stream, we allocate some heap memory to pass it several arguments. While this does work, it is a problem, for two reasons:

1. It goes against the design principle of these wrappers, which is being thin, and refraining from forcing the user into the use of particular data structures or programming patterns that the underlying API itself does not force.
2. More importantly - this will be incompatible with repeated launching of the same function, which is what happens if we apply stream capture, then get the result into an execution graph template.

We therefore need to avoid this allocation, and make do with only what CUDA's API itself guarantees the provision of.

And - note also the deprecation of the callback enqueuing API function which passes the stream to the callback; we'll probably need to make do without it.

[eyalroz]

Farwell fancy-shmancy closure call with arguments, we hardly knew ya!