Description: fix memory profile for partial graph run

Memory Profile is helpful for analyze memory peak used in static ORT planning.

This PR targets to fix the issues when enable memory profile in ORTModule partial graph run. Two issues:

1. Loose the failure if some tensors are missing in track.
2. Only summarized after backward partial graph run.

Validated with
./build.sh --config $flavor --use_cuda --enable_training --build_wheel --skip_tests --cuda_version=11.3 --parallel 8 --enable_training_torch_interop --cmake_extra_defines onnxruntime_ENABLE_MEMORY_PROFILE=ON

and

pytest -k test_forward_call_single_positional_argument orttraining/orttraining/test/python/orttraining_test_ortmodule_api.py

Motivation and Context

• Why is this change required? What problem does it solve?
• If it fixes an open issue, please link to the issue here.