Add bazel cache for release jobs

[Tombana]

What do these changes do?

This adds a bazel cache for release jobs:

• if they timeout we can simply restart them and they should finish within the time limit
• if they fail for any other reason we can try fixes without having to wait for 6 hours

How Has This Been Tested?

It is running here: https://github.com/larq/compute-engine/actions/runs/2200323727

[Tombana]

New run with different folders for each python version, and with passing the credentials into the manylinux container: https://github.com/larq/compute-engine/actions/runs/2200491384

[lgeiger]

Awesome! I just have one minor comment, other than that this looks great!

[lgeiger]

For the release workflow this should always exist, but I agree it's probably good to keep this check just to be safe.

[Tombana]

On the previous run only the python 3.7 job failed with an internal gcc error.
I've applied Lukas' suggestion of mounting the gcloud credentials directly, and I've set off another run: https://github.com/larq/compute-engine/actions/runs/2202486415
If the cache is working as expected these should now all finish quickly.

[Tombana]

The cache is working but not completely: the builds take about ~25 minutes, so something is working but it's still rebuilding some parts.
(Some of the failing workflows are unrelated, caused by a failed download)

[CNugteren]

25 minutes should be good enough. It saves a lot of time, and it makes the Windows builds pass. However I still see some builds ongoing at the moment, so some might take longer?

[Tombana]

25 minutes should be good enough. It saves a lot of time, and it makes the Windows builds pass. However I still see some builds ongoing at the moment, so some might take longer?

Indeed. I don't really know why the cache is only partially working, but I think this PR can be merged nonetheless, as it already helps.