Fine performance metrics: client context manager

[crusaderky]

• Part of Fine performance metrics meta-issue #7665
• Blocked by Fine performance metrics: Store data on the Scheduler #7666
• Soft-blocked by Fine performance metrics: Break down idle time on the Worker #7671
• Soft-blocked by Fine performance metrics: Break down idle time on the Scheduler #7672

Add a trivial context manager in the client to measure a run end to end.
Rough POC below; the real thing would need to work with async clients and use RPC instead of run_on_scheduler:

@contextmanager
def get_run_metrics(self) -> Iterator[dict[tuple[str, ...], float]]:
    t0 = perf_counter()
    before = client.run_on_scheduler(lambda dask_scheduler: dask_scheduler.cumulative_worker_metrics)
    out = {}
    yield out
    after = client.run_on_scheduler(lambda dask_scheduler: dask_scheduler.cumulative_worker_metrics)
    t1 = perf_counter()
    for k, v in before.items:
        after[k] -= v
    out.update(after)
    observed_time =  (t1 - t0) * cluster_threads 
    execute_time = sum(
        v for k, v in after.items() 
        if isinstance(k, tuple) and k[0] == "execute" and k[-1] == "seconds"
    )
    out["execute", "n/a", "client", "seconds"] = max(0, observed_time - execute_time)


with c.get_run_metrics() as metrics:
    my_collection.compute()

The actual method name could use improvement.

Note that the above snippet allows measuring the time spent in client->scheduler comms - which can be very substantial.
It will be very large unless we also deliver #7671 and #7672.

[crusaderky]

The above design needs to be overhauled on top of the new span context manager

• XREF User-defined spans #7860