MAINT: Publish the backend log from the shared test jobs

[varunj-msft]

Description

The end-to-end and integration jobs launch pyrit_backend through ServerLauncher, which writes the server log to a file in the system temp directory and then throws it away when the agent is torn down. A failing job therefore reports only the client side of the failure, which means a backend that is alive but unresponsive is indistinguishable from one that never started at all.

That distinction is exactly what we could not make while debugging the current End to End Tests failures: every one of the failures is a client-side read timeout, and the server's own account of what it was doing is discarded along with the agent.

This collects the log and publishes it as a pipeline artifact, following the pattern adversarial-benchmark.yml already uses:

• The collection step runs under condition: always(), so a job that fails, or times out at the 360-minute cap, still yields the log.
• ServerLauncher writes to tempfile.gettempdir(), which honours TMPDIR/TEMP/TMP before falling back to /tmp, so the step checks the agent temp directory as well as /tmp.
• A missing file is not an error. Jobs that never launch a backend log one line and move on.
• The publish task is skipped unless the copy actually landed, so we never publish an empty artifact that claims a log is there when it isn't.

Only .azuredevops/test-job-template.yml changes. No product code, no test code, and no behaviour change for any job — this adds diagnostics only.

Part of the v1.1.0 release wave with #2510, #2511 and #2512.

Tests and Documentation

This is pipeline YAML, so no unit test can cover it. I validated it by extracting the script body and executing it against all three paths:

• Log present in the temp directory: copied, backendLogFound=true, artifact published. Step exits 0.
• No log present, i.e. a job that never starts a backend: prints "No pyrit_backend.log found", publish skipped. Step exits 0.
• Log present but the copy fails, tested with an unwritable staging directory: flag not set, publish skipped. Step still exits 0.

The third case is worth calling out, because the first version of this step got it wrong. set -uo pipefail does not include -e, so a failing cp is silent: the step would have claimed success, set backendLogFound=true, and published an empty directory while simultaneously printing "No pyrit_backend.log found". The step now gates both the success message and the setvariable on the copy actually succeeding, so the flag is set only when the file has really landed.

No documentation changes, since this touches only pipeline YAML. JupyText was not run and is not applicable here: no notebooks, docs or public APIs are affected by this change.