Add parallel file processing with rayon

[Zious11]

Summary

Process multiple PCAP files concurrently using rayon for significant speedup on multi-file inputs.

Requirements

• Use rayon (already in dependencies) to parallelize file-level processing
• Each file gets its own Summary + analyzer instances
• Merge per-file results after parallel processing completes
• Add Summary::merge(&mut self, other: Summary) method
• Maintain deterministic output ordering (sort by filename)
• Progress bar should show overall progress across all files

Acceptance Criteria

• rayon::par_iter() over resolved file list
• Per-file analysis with thread-local Summary/analyzers
• Merge step that combines results
• Tests verifying merge produces correct totals
• Benchmark showing speedup on multi-file input

[Zious11]

Triage 2026-07-08 (codebase-validated at develop c4eb1f4; research-validated per DF-VALIDATION-001): Still valid, deferred to a perf-driven wave. Premise correction: rayon is NOT currently in [dependencies] — it appears only in Cargo.lock as a transitive dev-dependency via criterion; production code has no access to it. External validation (2026-07-08): rayon 1.11.0, MSRV 1.80 (compatible), no RUSTSEC/OSV advisories; per-file parallelism is the industry-standard shape for offline PCAP CLIs (Zeek/Suricata/tshark all process per-file offline). Pickup prerequisites: (a) add rayon = "1" as a runtime dep; (b) design Summary::merge(&mut self, other: Summary) (src/summary.rs has no merge API; HashMap fields merge additively); (c) parallelize the sequential file loops at src/main.rs:363 (run_analyze) and src/main.rs:801 (run_summary); (d) explicit AC for deterministic finding ordering under parallelism (per-file ordering key); (e) ADR-0004 atomics: read counters only post-fold, and AUDIT for cache-line false sharing — wrap hot adjacent counters in crossbeam_utils::CachePadded or #[repr(align(64))] before enabling parallel loops; (f) throttle indicatif progress updates (ProgressDrawTarget::stdout_with_hz(...) or coalesced increments) so the draw lock doesn't cap scaling; MultiProgress has no known deadlock with rayon but per-item inc(1) from hot loops causes redraw contention. Estimated size M (~5 pts). ADR-0002 purity boundary preserved (no cross-file flow reassembly exists).