New-pipeline: Memory management on PTX

After reference counting is made explicit (#569), we should modify the PTX backend to make use of the explicit reference counting instructions.

Note that the reference counting should be executed asynchronously, in the same way kernels are launched asynchronously on PTX.