feat: system benchmark script with redis comparison

[kacy]

summary

• adds bench/compare-redis.sh which uses redis-benchmark to measure end-to-end throughput (TCP + RESP3 + shard routing)
• tests SET and GET at 3B, 64B, and 1KB value sizes, both pipelined (P=16) and non-pipelined (P=1)
• supports --ember-only mode (no redis needed) and --json output for CI
• saves raw CSV results to bench/results/ with timestamps
• adds bench-compare and bench-quick Makefile targets

tested

ran bash bench/compare-redis.sh --ember-only on macOS, results:

test                      ember (rps)
SET (3B, P=16)              1,123,595
GET (3B, P=16)              1,010,101
SET (64B, P=16)             1,000,000
GET (64B, P=16)               952,381
SET (1KB, P=16)               602,409
GET (1KB, P=16)               869,565
SET (3B, P=1)                 136,798
GET (3B, P=1)                 137,362

design considerations

• uses redis-benchmark because it's the standard tool and already speaks RESP — no custom load generator needed
• starts and stops servers automatically with trap-based cleanup on any exit path
• pre-populates keys before GET benchmarks so they hit existing data
• environment variables allow tuning without modifying the script
• CSV output enables historical tracking and regression detection