Fixes #689: Completed removal of cuda::memory::async:

* Operations affected include: `allocate()`, `free()`, typed_set()`, `set()`, zero()`
* Combined (kinda-)synchronous and asynchronous operations using an `optional_ref<stream_t>` extra parameter, defauling to `{}`/`nullopt` for the synchronous version
* Already done with `copy()` in an earlier commit
* `prefetch()` and `prefetch_to_host()` also affected, but no additional parameter is necesary
* Operations affected include: `allocate()`, `free()`, typed_set()`, `set()`, zero()`
* Comment tweaks
* Removed some redundant code