rancher/rancher#37639

• Added retry flags for backups and s3, so that rke can control the number of retries. This fixes scenarios where rancher will continuously retry (previously 5 times), and rke-tools would retry separately (4 times per snapshot), which would end up choking rancher, rke, the downstream etcd nodes, and the s3 endpoint.
• Check whether the s3 bucket has versioning enabled, as in this case subsequent uploads to the s3 bucket would be useless. Although this versioning is not useful from a rancher perspective (since during the first restore we download the latest version and plow over the existing snapshot on all the etcd nodes), it does allow us to save on network traffic which has been the source of other issues while trying to restore from snapshots taken during this sort of network saturation, because taking snapshots (including uploading to the s3 bucket) is done node by node.
• Fixed an issue where failure to create an s3 client (such as a bad token, or the bucket doesn't exist) would cause the snapshot on the node to be deleted. Failure to upload a snapshot to s3 should not constitute a cluster with no backups ever.
• Removed using the v1 s3 API when the endpoint is storage.googleapis.com. I tested this with a bucket in GCP, and it worked fine with the v2 s3 API, so whatever requirement drove that is no longer valid.