Fix a DDP graph capture issue

[yaochengji]

There're issues in torch/xla DDP. #8396. After investigation, one of the issues are caused by the incorrect graph capture of the combination of inplace-update and non aten op.

Changes:

1. Modify TryGetXlaTensor to make sure we can always have the most updated version of tensors
2. Modify the learning rate of ddp test from 1e-4 to 1e-1 as 1e-4 is too small to verify the DDP correctness
3. Remove the gradient_as_bucket_view = True requirement in all the docs and code
4. Add the test_inplace_update.py file, note that test_non_aten_op_after_partial_update cannot get passed before the TryGetXlaTensor patch

[yaochengji]

cc @qihqi @tengyifei

[jeffhataws]

@tengyifei will you enable TPU testing on this PR?

[tengyifei]

@yaochengji could you force push a commit? I just added a tpuci label but we need to retrigger the jobs to run TPU tests.

[tengyifei]

Thanks!

[yaochengji]

@tengyifei thanks for your review. Do I still need to force push as it is already merged?

[tengyifei]

oops, i merged too soon :) looks like tip-of-tree build is green, so nothing needed on your part now!

[jeffhataws]

@yaochengji will you add this change to #8455 for cherry-pick to 2.6? Let me know if I can help.

[yaochengji]

@jeffhataws it's added, thanks.