Make the tool agnostic for the 'failures' file location.

There was a solution to distinguish between the jobs that store the failures.csv file in the results directory and those that store it in a subdirectory (named with the unsharded job name). But a new structure has been found in mesa.

So, instead of introducing one more exception, the solution is to walk the paths looking for the file. If there is more than one failures file (as is the case in the newer situation found in mesa), concatenate them to prepare later the expectations update as if it is a single file.

---

Requires !78 (merged)