Treat `bench()` errors as first-class iteration results

[wfxr]

Problem

If a benchmark iteration returns Err(...), collectors currently only record it in error_dist and do not increment iteration counters or record latency. This can lead to optimistic throughput/success ratios, especially since examples commonly use ? inside bench().

Current behavior

• Success ratio only considers statuses aggregated into IterStats (src/report.rs:41-54)
• Errors are tracked separately and do not contribute to iters/items/bytes (src/collector/silent.rs:59-73, src/collector/tui.rs:200-225)

Proposed solution

Option A (compatible, recommended): move duration measurement into the runner so it can always attach a duration to both Ok and Err outcomes, and change the result channel to carry a unified iteration outcome:

pub enum IterOutcome {
    Ok(IterReport),
    Err { duration: Duration, message: String },
}

Collectors would then:

• Increment iteration counters for Err
• Record duration into the histogram
• Add a synthetic Status::error(...) entry into status_dist
• Keep error_dist for messages

Option B (docs-only): document the semantics clearly and update examples to convert errors to IterReport.status, but keep current behavior.

Acceptance criteria

• Iterations that fail with Err are reflected in total iteration counts
• Success ratio and throughput include failures in the denominator
• Error details remain available in error_dist
• Examples demonstrate the recommended error-handling pattern