Make Transforms smarter when device memory is not enough to fit in

We can make our transforms smarter to decide if input (image) fits into the available device.
Every device comes with a max memory size.

If I have an image which doesn't fit into the GPU then I would rather run it on CPU.
Some of transforms can help to determine if we need to fallback or can continue to run on GPU.

In many cases, users see OOM because at some transform it fails due to input not able to fit enough into the available device/GPU.
Then for that sample he either has to pick a new device with larger memory or have to disable gpu

We need an auto-fallback feature in very simple cases.
(As first step, I am even ok if do based on max memory size of device vs current available memory)

No files, tests, or entry points are named. Start by locating the transform implementations and the existing device-selection and memory-handling logic, then clarify which simple cases should fall back from GPU to CPU and how completion will be verified without out-of-memory failures.