[dattri.algorithm] Add TRAKAttributor and related examples

[TheaperDeng]

Description

1. Motivation and Context

Add TRAK as one of the attributor in our package

2. Summary of the change

1. Add TRAKAttributor with similar usage as other attributor. The only difference is that it requires a correct_possibility_func for the calculation of Q.
2. Add an example running on mnist_lr to find the mislabeled data and it performs well.

   Peak memory usage: 39.129088 MB                                                                                                                                                                      
   torch.Size([1000])
   [(0, 0), (100, 74), (200, 97), (300, 97), (400, 98), (500, 99), (600, 99), (700, 99), (800, 99), (900, 99)]
   Checked Data Sample      Found flipped Sample     
   --------------------------------------------------
   0                        0                        
   100                      74                       
   200                      97                       
   300                      97                       
   400                      98                       
   500                      99                       
   600                      99                       
   700                      99                       
   800                      99                       
   900                      99

3. What tests have been added/updated for the change?

• [ ] Unit test: Typically, this should be included if you implemented a new function/fixed a bug.

[tingwl0122]

will take a look later

[tingwl0122]

will take a look later

As mentioned in #85 , probably we can test TRAK performance on CIFAR-2 also? will need to add one more example script.

[tingwl0122]

Hi @TheaperDeng , overall, do you think the unsqueeze(0) style stuff within the f,m function is acceptable? Not sure whether we can get rid of that by some different vmap configuration.

[tingwl0122]

Hi @TheaperDeng , overall, do you think the unsqueeze(0) style stuff within the f,m function is acceptable? Not sure whether we can get rid of that by some different vmap configuration.

now leave this as a TO-DO, please refer to #86 .