Preserve soak evidence when resource sampling fails

[rmcdaniel]

Summary

Fixes the sampler crash observed in scheduled run 34597695626; related to #137.

• Accept MySQL counts only after a successful command with exactly four nonnegative integer fields. Nonnumeric Unavailable: output becomes a failed observation, not ValueError.
• Record OS/timeout failures per collector and continue the remaining collectors and later samples. Failed samples still disqualify trusted evidence; missing counts are not reported as zero.
• Add regression coverage for malformed/error output, valid zero counts, each collector failing, and recovery on the next sample.

Validation

python3 -m unittest tests/Unit/Support/server_soak_test.py tests/Unit/Support/vultr_soak_cleanup_test.py: 27 tests pass. py_compile, diff check and repository publication check pass.

The daily schedule is paused (DW_PERF_SOAK_ENABLED=false) pending repair and review, with no extra paid run triggered. The prior run also recorded six workflow result timeouts and two readiness timeouts; this instrumentation change does not fix or waive those failures and is not an endurance/capacity pass.

No public runtime/protocol/default-image behavior changes. No package release or SDK rebuild is required for this harness-only change; normal public CI owns the smoke check.