onnxruntime

Zhijxu/cleanup cached tensors when oom (#19306)

in pytorch, when oom happens at bp, user could decrease the batch size
and rerun it without restarting the process.

while in ORT, the intermediate tensors are kept even OOM, so decrease
batch size still fail.


this is torch run, we can see after oom failure, torch will release
tensor before next step

![image](https://github.com/microsoft/onnxruntime/assets/43435212/92b8a2e3-454b-448a-a223-17cb91d463c2)

this is from ort, we can see ort not release its tensors after OOM
failure.

![image](https://github.com/microsoft/onnxruntime/assets/43435212/bb6a3882-8e14-4f37-8079-e7f70fc2546b)

ort with the PR, we can see memory is released, **the 4GB memory is not
own by ort, and will be released by torch at the end**.

![image](https://github.com/microsoft/onnxruntime/assets/43435212/7f39d711-4e36-47d5-aecf-3805433a6d01)