Fix GCP Batch co-execution deadlocks

[ksuderman]

Part of #465.

This PR ensures the tool and sidecar runnables are executed in the correct order and don't deadlock.

This PR also:

1. Ensures outputs from intermediate jobs are collected.
   Galaxy's path rewriting causes tools to write outputs into working/ rather than outputs/, but __collect_outputs() only checked the outputs directory, so those files were skipped. Also relaxes the dynamic collection pattern from dataset_\d+\.dat to dataset_[\w-]+\.dat so UUID-named datasets (e.g. dataset_b0e57817-…​.dat) match.
2. Re-raises infrastructure errors so they fail the job rather than producing a "successful" job with empty outputs. FileNotFoundError is deliberately excluded.
3. Pass docker_extra_volumes to the Batch containers
   Comma-separated Docker -v-style volumes (used to bind-mount CVMFS paths such as
   /cvmfs/data.galaxyproject.org) are now applied to both the sidecar and tool runnable
   containers, so tools that depend on CVMFS reference data work.

[mvdbeek]

Could you pull out d2487a8 and 7244d2c into separate PRs ? I'm happy to merge the rest

[ksuderman]

Could you pull out d2487a8 and 7244d2c into separate PRs ? I'm happy to merge the rest

I will open another PR with these two commits in a moment.