A probe whose condition threw was silently treated as not matching, so
a broken condition (a typo in a variable name, a member access on an
undefined value, ...) left the user with a probe that never fires and no
explanation. The condition was also evaluated again, and threw again, on
every single hit.

Condition errors are now handled like the other tracers do it: the probe
result carries the error in `evaluationErrors` (with the condition DSL
as `expr` and the error as `message`) and nothing is captured, so the
user can see why the probe is not producing regular results. Error
results bypass the per-probe and global rate limits, and are instead
limited by a per-probe throttle: after an error the condition is not
evaluated again for five minutes, and one error result is reported per
window. This also bounds the cost of a condition that is expensive to
fail.

The probe sampler records the error on the main thread and flags the
sampled probe index, so the worker knows to fetch the error from the
paused frame instead of capturing a snapshot.