Allow using tag prefixes in calculate-docker-image

[atalman]

Issue with discrepancy between dockerhub image names for manylinux 2_28 images and ecr images:

Current naming:
308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/manylinux2_28-builder-rocm6.2.4:804413e4d6f4e378061d47bd4ed99c1cbb462011

Proposed naming:
308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/manylinux2_28-builder:rocm6.2.4-804413e4d6f4e378061d47bd4ed99c1cbb462011
similar to
pytorch/manylinux2_28-builder:rocm6.2.4-eed9bb3a0e85082d4d95ad4fa33673a158b8c488
This will rid us of the need to request new ECR entries to be created in the internal docker registry every time there's a ROCm (or even CUDA) version update, and afaict, there'll be no loss of information.
Also, it'll sync up the two repos' naming conventions perfectly.

[jithunnair-amd]

Thanks for considering my suggestion, @atalman.

Also related: it'd be really helpful to be able to test building binaries against the internal docker registry docker images created by a PR's changes, without needing to merge the PR. Today, this is not straightforward because the binary build jobs use dockerhub images e.g: https://github.com/pytorch/pytorch/blob/fda43c98d1bc82eeb493f6e79a1fd5f1636474a5/.github/workflows/generated-linux-binary-manywheel-nightly.yml#L4
But esp. if we do the above suggested naming alignment, it could be as simple as switching between internal registry and dockerhub based on some notion of is_pr_build?

In fact, the PR binary workflow runs should always use the internal ECR docker images instead of the dockerhub ones. That way, if a PR modifies the logic for building the manylinux docker images, the binaries will test those changes, otherwise it'll just pick up the already-existing docker image which should be practically identical to the one in dockerhub anyway, due to matching sha tag.

[ZainRizvi]

Instead of moving the tags over, can we strip the version numbers from the ECR name and just use perhaps two ECRs, one for vCurrent and one for vNext?

[clee2000]

Some combination of

• Action for building docker binary builds pytorch#151471
• Use reusable binary docker build action for almalinux, clean up script pytorch#151483
• Use reusable binary docker build action for libtorch pytorch#151488
• Use reusable binary docker build action for manywheel pytorch#151489
• Use folder tagged docker images for binary builds pytorch#151706
• Custom script path and tag prefix for calculate docker image #6499

should fix everything listed here:

• difference between docker io names and image names built on PR that get pushed to ecr. Images built on PR now follow the docker io images of <imagename>:<cuda, cpu etc info>-<additional branch info>
• due to the above, we no longer need to manually create a new ecr when a new cuda/rocm version is added
• adding ciflow/binaries on a PR will now use the image that gets built on the PR, assuming that the PR does need to build a new docker image

so I'm going to close this as completed. Please re open if I missed something