Add shadow workflow for hi-ml-pr

[mebristo]

Currently, status checks don't run on the PR pipelines because they involve path filtering. This PR follows the suggestion in github documentation to add a second workflow that runs on the inverse of the filtered paths, to enable status checks to run

[mebristo]

Not really ready for review, but can't see whether merging would be blocked until the PR is marked as ready

[codecov[bot]]

Codecov Report

Merging #595 (3b77f08) into main (9da88a6) will increase coverage by 9.69%. The diff coverage is n/a.

Flag	Coverage Δ
hi-ml	81.92% <ø> (ø)
hi-ml-azure	25.17% <ø> (ø)
hi-ml-cpath	79.14% <ø> (+22.50%)	:arrow_up:
hi-ml-multimodal	86.54% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
hi-ml-cpath/src/SSL/data/cxr_datasets.py	62.88% <0.00%> (ø)
hi-ml-cpath/src/SSL/ssl_augmentation_config.py	100.00% <0.00%> (ø)
...l-cpath/src/SSL/lightning_modules/simclr_module.py	100.00% <0.00%> (ø)
hi-ml-cpath/src/SSL/data/image_transforms.py	92.59% <0.00%> (ø)
...c/SSL/lightning_containers/ssl_image_classifier.py	93.10% <0.00%> (ø)
hi-ml-cpath/src/SSL/utils.py	80.30% <0.00%> (ø)
hi-ml-cpath/src/SSL/data/io_util.py	75.86% <0.00%> (ø)
...path/src/SSL/lightning_modules/byol/byol_module.py	97.18% <0.00%> (ø)
hi-ml-cpath/src/SSL/data/datamodules.py	94.66% <0.00%> (ø)
...path/src/SSL/lightning_modules/byol/byol_models.py	100.00% <0.00%> (ø)
... and 39 more

[ant0nsc]

There's one other thing we should probably do in this PR: Give all jobs a unique name. We have, for example, 3 jobs that are called flake8, one for each project. When adding flake8 as a necessary check, which flake8 are we getting? So could rename them the flake8-hi-ml, flake8-cpath, flake8-multimodal. Or hi-ml-flake8, cpath-flake8, ... - Not sure which of the two is better, but unique names for sure.

[ant0nsc]

Just discussed with @peterhessey that we could simplify the shadow workflows by only running the "final" step of each PR workflow, like publish testpypi for hi-ml. This way, we don't have to replicate (and maintain) the full complex graph of each PR, only need the final step that depends on the success of all previous ones.