Instead of suppressing calls to free after closing the cuda context, delay closing the cuda context until all calls to free have been made. Might not be *quite* as performant, but feels perhaps a little cleaner.