[runtime] enable runtime stitching

[pilkicTT]

The tt-mlir runtime now leaves every tensor on device and can accept inputs already on device. Use that to avoid unnecessary movements of tensors to and from the device.

For example, when running inference of a model in a loop, we are currently sending the weights/constant tensors on each iteration. This is obviosly not ideal.

Another example would be in training scenario. We can now reuse outputs of, let's say, forward() program on device as inputs to the backward() program. Without needing to move outputs of forward() to the host and then running the backward() program which will move them back to the device.