Add PTQ example for PyTorch CV - Segment Anything Model

[cfgfung]

Type of Change

I am adding one more CV example (first segmentation foundation model) to demonstrate the usage of INC.

Description

SAM is the first segmentation foundation model in the CV domain. It is composed of ViT and several image decoding layers. By applying suitable quantizations on specific layers, it is found that there is only a slight drop in the dice metric (~0.01). This can be an example of CV users adapting INC in their use case.

Expected Behavior & Potential Risk

Users will base on the example code to experiment INC on SAM. No risk to the existing code base.

How has this PR been tested?

Replicate the steps according to the readme in a new environment.
It is tested on Intel(R) Xeon(R) Platinum 8280 CPU @ 2.70GHz

Dependency Change?

Users have to install the following packages in order to run the example. 'torchmetrics' is a new package to the user.

torchvision
tdqm
torchmetrics
neural-compressor

[violetch24]

Please refer to torchvision example for a similar format and process:

1. After quantization, q_model needs to be saved, then loaded for accuracy and benchmark.
2. Users should be able to load either fp32 or int8 model.
3. Need to add benchmark part using INC's BenchmarkConfig.
4. The example should include bash scripts to run tuning and benchmark separately.

[cfgfung]

Hi,
I have aligned the codes and tested them on Endeavour (SPR node).
Would you please help to see if this is ready for PR?

[violetch24]

Please update the flag name in comment as well as in shell scripts and README. There are also some spell errors in pre-ci. Others look good to me, thanks!

[chensuyue]

There are 2 new components used by this PR, segment-anything and torchmetrics, I will add into the BoM list.
And in order to track this example with our weekly test, I will submit another PR to unify the test interface.