Allocate cuda stream `let_value`'s `__op_state_` in pinned memory

[ericniebler]

The fix for #953 is not strictly correct. Only let_value's operation state needs to be in pinned memory, and it needs to be there not just for sync_wait but for all algorithms that eagerly connect a sender and store the resulting operation state in memory that isn't device-accessible.

Once let_value's operation state is allocate in pinned memory, back out the change to sync_wait that was made in #970.