[Feature] Possibility to pass non_blocking attribute to convert_tensor #231
Description
Following torch docs on CUDA memory management if user configures DataLoader with pin_memory=True then
he/she can use asynchronous GPU copies, just passing an additional non_blocking=True argument to a cuda() call or tensor method .to().
In ignite this can be done with utils method convert_tensor that accepts device argument: cpu, cuda or cuda:N. It would be great to have a possibility to pass non_blocking as kwargs.
In addition, this should be also propaged until create_supervised_trainer and _prepare_batch.
PS: it is okay to call x.to('cpu', non_blocking=True), no error raised.