Fixed on_disk_cache issues

[VitalyFedyunin]

Fixed issues with cache locks and cache files overwrites. Required to be compatible with pytorch/data#810

Stack from ghstack (oldest at bottom):

• -> Fixed on_disk_cache issues #1942

[Nayef211]

@VitalyFedyunin looks like all TestCNNDM is failing on all platforms. Will stamp once those are fixed!

[VitalyFedyunin]

Looks like tests havn't picked up recent torchdata changes, lets wait a bit and restart them.

[joecummings]

Looks like tests havn't picked up recent torchdata changes, lets wait a bit and restart them.

These tests are now failing for every new PR. Which specific change is breaking this? @VitalyFedyunin

[VitalyFedyunin]

That PR pytorch/data#810 introduced BC breaking change on cache api (as well as major stability fix).
This PR is fixing all failed tests.

[VitalyFedyunin]

Not sure why 3.8 is failing. Something related to xml decoding.

[joecummings]

Not sure why 3.8 is failing. Something related to xml decoding.

I think this is a transient error with that particular dataset. I'll approve but @Nayef211 can you confirm?

[joecummings]

lgtm

[Nayef211]

Not sure why 3.8 is failing. Something related to xml decoding.

I think this is a transient error with that particular dataset. I'll approve but @Nayef211 can you confirm?

Yup can confirm. This was also raised in #1926

[ejguan]

@Nayef211 @abhinavarora Can we cherry pick this PR to the release?
We plan to cherry pick the one from torchdata pytorch/data#810. And, since this is the PR fixing a bug #1903, should this be included in the release?

Edit: If you can cherrypick this PR, I will do anther RC cut from torchdata side and upload a new RC release for torchtext to validate the change on release branch.