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

[acere]

Closes #55

What

Adds low-memory mode for large-scale test runs, replaces _builtin_stats with incremental RunningStats, and shows live stats on the progress bar during runs.

Changes

llmeter/utils.py

• New RunningStats class that accumulates metrics incrementally via sorted value lists. Provides:
  • update() — records one response's metrics
  • to_stats() — computes all stats as raw numeric values (single source of truth)
  • snapshot() — formats a configurable subset for tqdm progress-bar display
  • DEFAULT_SNAPSHOT_STATS — default fields: p50/p90 TTFT, p50/p90 TTLT, median tokens/s, total input/output tokens, failure count

llmeter/runner.py

• New low_memory parameter on _RunConfig/Runner/run(): writes responses to disk without keeping them in memory. Requires output_path.
• New progress_bar_stats parameter: configurable live stats on the progress bar.
• RunningStats is always created and fed during _process_results_from_q. In low-memory mode, responses are not appended to _responses.
• _run() always populates result._preloaded_stats from RunningStats.to_stats().

llmeter/results.py

• Removed _builtin_stats @cached_property (and functools.cached_property import).
• New _compute_stats() classmethod as fallback for manually constructed Results.
• stats property now reads from _preloaded_stats first, falls back to _compute_stats().
• load_responses() recomputes _preloaded_stats from loaded data.
• Result.load(load_responses=True) computes _preloaded_stats from responses.

Tests

• Updated test_lazy_load.py: replaced _builtin_stats cache invalidation test with _preloaded_stats recomputation test.
• Updated test_results.py: replaced _builtin_stats caching assertion.

Usage

# Low-memory mode
result = await runner.run(output_path="/tmp/run", low_memory=True)
result.stats              # works (computed incrementally)
result.responses          # [] (not in memory)
result.load_responses()   # loads from disk

# Custom progress-bar stats
result = await runner.run(
    progress_bar_stats={
        "p99_ttlt": ("time_to_last_token", "p99"),
        "tps": ("time_per_output_token", "p50", "inv"),
        "fail": "failed",
    },
)