Add TorchData as hard dependency

[ejguan]

Per title

• Add install_torchdata.sh as the pre-script to install torchdata and specify TORCHDATA_VERSION and CONDA_TORCHDATA_CONSTRAINT in BUILD_ENV_FILE

[ejguan]

I added TORCHDATA_VERSION against all the places with PYTORCH_VERSION. But, I am not sure how build is going to use it for torchtext releases.

[ejguan]

@Nayef211 @osalpekar
I did some research on the build system when I tried to enable torchdata as the hard dependency. I think there are a few requirements from test-infra to make torchdata as a hard dependency. When building torchtext, torch is installed from link. If I want to install torchdata via pre-script, due to pre-script only allows a file path, I don't have a way to know if the channel is nightly or test, and I have to add different scripts for conda or wheel and hardcoded the pre-script per build workflow. Besides, I am not able to specify the version of torchdata during the step of building torchtext binary.

In order to resolve this problem, I personally think it would be good to run source "${BUILD_ENV_FILE}" in the step of pre-script.

[osalpekar]

@ejguan Thanks for the investigation - we can certainly source ${BUILD_ENV_FILE} before running the pre/post scripts. I'll try to put up a PR for this later today.

[osalpekar]

@ejguan PR is up: pytorch/test-infra#1073. Will monitor builds to ensure no existing builds are broken, but otherwise feel free to review/experiment with.

[ejguan]

I think the failed test is unrelated to this PR. Edit: All tests passed

[Nayef211]

LGTM. Thanks so much for taking care of adding the torchdata dependency @ejguan!