fix(bench): stop recording probe errors as tool versions

[bryantbiggs]

check_tool never inspected the exit status and fell back to stderr, so a tool that exits non-zero from its version probe had the error text stored as its version. dregsy has no version flag at all, so 10 of its 13 records in bench/results/ecr.json hold a Go flag error where a version belongs:

"version": "time=\"2026-04-26T13:17:46Z\" level=error msg=\"flag provided but not defined: -version\""

That defeats what the archive is tracked in git for. A run record that cannot say which dregsy produced a number cannot attribute a shift to a version change.

Tolerating a failed probe is specific to the tool that has no version to report. check_tool is the only gate that runs before a benchmark starts, so for a tool that does report one, a failed probe is now an error rather than a quietly absent version -- otherwise a broken ocync or regsync would sail through the pre-flight check looking fine. reports_version matches exhaustively with no wildcard, so adding a tool forces that decision rather than defaulting to lenient.

Version parsing moves into a pure function so both paths are testable without spawning a process. The two guards were kill-tested rather than assumed: removing the exit-status check fails failed_probe_is_not_recorded_as_a_version and nothing else, and making every tool exempt fails failed_probe_errors_for_a_tool_that_reports_a_version and nothing else. The dregsy fixture is the exact stderr string the archive actually stored, not an invented one.

check_tool was the only site in xtask ignoring exit status -- report.rs uses .filter(|o| o.status.success()) and remote.rs checks explicitly -- so this makes it consistent with the surrounding code rather than introducing a new pattern.

Two related gaps are left alone, both pre-existing and neither caused here. check_tool spawns the tool by name from PATH while run_tool runs ocync from target/release/ocync, and it runs before build_ocync, so the recorded ocync version is not necessarily the binary that was benchmarked. Separately, skopeo performs every transfer dregsy is credited with but is not a Tool variant, so its version reaches no record. Both need the bench flow reordered or the record schema extended, which is a larger change than this one.