We need to think about temporary vs. permanent failures in source mirroring

[See original issue on GitLab](https://gitlab.com/BuildStream/buildstream/-/issues/522)
In GitLab by [[Gitlab user @jonathanmaw]](https://gitlab.com/jonathanmaw) on Jul 27, 2018, 13:38

## Background

[//]: # (Provide a background or the root/source that justifies this task or action.)

The Source Mirroring implementation currently treats any kind of BstError raised in `Source.fetch()`, `Source.track()` or SourceFetcher.fetch()` as non-critical, and only if all fetch/track calls fail does it raise the last exception.
This is going to have unexpected behaviour when the final call was a temporary error, but the previous ones were not.

## Task description

[//]: # (Short summary of the action to be executed)

* Check whether the below plan makes sense.
* When the error is temporary, try the next call.
* When the error is not temporary, raise it.

## Acceptance Criteria

[//]: # (Acceptance criteria should follow the S.M.A.R.T. principle https://en.wikipedia.org/wiki/SMART_criteria )

* When a non-temporary error is raised by `Source.fetch()`, `Source.track()` or `SourceFetcher.fetch()`, buildstream aborts.
* When a temporary error is raised by the above methods, but not every call, buildstream proceeds as normal.
* When a temporary error is raised every time the above methods are called, buildstream aborts.

----

Start by tracing source-mirroring calls to Source.fetch(), Source.track(), and SourceFetcher.fetch(), then inspect how BstError failures are classified and propagated. Verify the behavior against the acceptance criteria: non-temporary errors abort immediately, temporary errors allow later calls, and failure of every call still aborts.

Written by the indexing model from the issue text.