Add security benchmark suite and performance overhead measurements

[vgudur-dev]

Description

Create a benchmark suite that measures:

1. Detection accuracy against known memory poisoning attack patterns
2. False positive rates on benign memory operations
3. Performance overhead (latency added per read/write operation)

Motivation

Enterprise adoption requires quantified security guarantees and performance impact data. This benchmark will:

• Provide data for the README and documentation
• Enable regression testing as new detectors are added
• Give users confidence in production deployment

Proposed Implementation

Create benchmarks/ directory with:

1. benchmarks/security_accuracy.py

• Test against a corpus of known attack patterns (prompt injection variants, obfuscated secrets, etc.)
• Measure true positive rate, false positive rate, and false negative rate
• Report results as a confusion matrix

2. benchmarks/performance.py

• Measure latency overhead per operation (read, write, snapshot, rollback)
• Test with varying policy complexity (1 rule vs 10 rules)
• Test with varying value sizes (100B, 1KB, 10KB, 100KB)
• Compare InMemoryStore vs Redis backend

3. benchmarks/attack_corpus/

• Collection of memory poisoning payloads for testing
• Categorized by attack type (injection, leakage, tampering, churn)

Acceptance Criteria

• Security accuracy benchmark with reproducible results
• Performance benchmark with latency measurements
• Attack corpus with 50+ test payloads
• Results table in README showing detection rates
• CI job that runs benchmarks on each release