QFw: write a job's story in five log lines, stitched to its trace

[DougSO]

Follow-up to #123 from wiring Loki into the cluster: the fake IQM logs nothing of its own inside a job's spans (its two per-job lines belong to the device query the client makes before the job starts), so its traces had no lines under the waterfall.

With the logs tier on, the instrumentation layer now writes the job's milestones itself, at debug like the rest of what QFw logs about a job:

• client: submitted job <id>: N circuit(s), S shots and job <id> completed after T ms;
• QPM: received circuit <cid> (qtask N): Q qubits, S shots, executing circuit <cid> on <device> via <api path>, circuit <cid> completed on <device> after T ms.

Each line is written while the job's span is current, so the tier stamps it with the trace and span ids. Two of those spans are not current on their own, the execute span begin_backend_execution opens and the job span outside job_scope, so the helper makes the span current for the one call. With the tier off the lines cost a boolean test and write nothing, to Loki or to DEFw's file.

The mock suite now runs the tier at debug and asserts the five lines against the trace ids of the QPM run and the client job. Syntax clean, mock 587 passed, unit 124 passed.

[DougSO]

Second commit, found by running this on the cluster: a DEFw service sets its Python log level after QFw's configure(), and DEFw's set_logging_level_helper removes every root handler when it does, the tier's included. In every real service process the tier therefore believed it was on while nothing was listening; the suite, without DEFw, could not see it. logs_enabled() now re-attaches the handler when it finds it gone, the story lines ask it before writing, and a unit test strips the root logger the way DEFw does and checks the recovery. The README's caveat is reworded accordingly.

[DougSO]

Rebased onto main at c5a3048 and stacked on #124, since both change the same telemetry files: this branch now carries #124's commit followed by its own two, so the two PRs no longer collide. Merge #124 first and this one merges clean afterwards. (If this one were merged first, #124 would show no changes and could simply be closed.) One conflict on the rebase, in the design doc's logs bullet; resolved by keeping #124's wording and appending this PR's sentence. Syntax clean, mock 619 passed, unit 126 passed on the stacked tree. The earlier ci-mock failure was test_foreground_run_stops_on_sigterm, a service-plane test with a five-second readiness wait that passes here three times in a row and is unrelated to these changes.