reset read offset when download blob failed

[Kila2]

Reset the read offset when downloading a blob fails to address the issue of some remote cache servers not supporting slice downloads.

[werkt]

This needs to be made into an option, and disabled by default. Without a progressive read, you are limited to the available bandwidth to server / concurrent downloads * remote_timeout. There is no reason to make this global.

Or you can do the nicer bit and recognize an InvalidArgument or some such return and then reset.

[Kila2]

Thank you for the suggestion. After reviewing, it seems challenging to implement this efficiently without introducing unnecessary complexity. Simply resetting CountingOutputStream is not ideal. If possible, we could consider retaining the read offset for DEADLINE_EXCEEDED status and resetting it for other statuses. For now, I’ll close this PR but appreciate your feedback and hope to revisit it with a better approach in the future.