feat: low-memory mode and live progress-bar stats

[acere]

Summary

For large-scale test runs, keeping all InvocationResponse objects in memory can be expensive — the bulk of the cost comes from response_text, input_payload, and input_prompt fields.

This issue proposes two related features:

1. Low-memory mode (low_memory=True)

A new low_memory parameter on Runner/run() that writes responses to disk as they arrive but does not accumulate them in the in-memory _responses list. Stats are computed incrementally. Requires output_path to be set. Responses can be loaded on demand via result.load_responses().

result = await runner.run(output_path="/tmp/my_run", low_memory=True)
result.stats          # works (computed incrementally)
result.responses      # [] (empty — not in memory)
result.load_responses()  # loads from disk when needed

2. RunningStats accumulator

A new RunningStats class that tracks metrics incrementally (counts, sums, sorted value lists for percentile computation). This replaces the _builtin_stats @cached_property on Result — stats are now always computed during the run and stored as _preloaded_stats, eliminating redundant recomputation.

3. Live progress-bar stats

RunningStats.snapshot() formats a configurable subset of live stats for tqdm display during the run: p50/p90 TTFT, p50/p90 TTLT, median output tokens/s, total input/output tokens, and failure count. Configurable via the progress_bar_stats parameter.

result = await runner.run(
    progress_bar_stats={
        "p99_ttlt": ("time_to_last_token", "p99"),
        "tps": ("time_per_output_token", "p50", "inv"),
        "fail": "failed",
    },
)