feat(bench): implement baseline comparison (--baseline flag)

[bug-ops]

Description

Run the same benchmark twice — once with [memory] enabled=true and once with [memory] enabled=false — and produce a delta comparison showing Zeph's memory value.

Part of epic #2827. See spec: .local/specs/zeph-bench/spec.md FR-006, US-004.

Scope

• --baseline flag on zeph bench run
• Runs the full scenario set twice: first pass with memory enabled, second pass with memory disabled (config override)
• Writes baseline/memory-on/ and baseline/memory-off/ result directories
• Top-level summary.md includes a delta table: per-scenario delta score, aggregate delta, interpretation note
• BaselineComparison struct serialized to top-level comparison.json

Acceptance Criteria

• Both memory-on and memory-off result files written to correct subdirectories
• Delta table present in top-level summary.md
• Aggregate delta = mean(memory-on scores) - mean(memory-off scores)
• Each pass uses the same isolation reset between scenarios