performance: implement adaptive worker scaling based on runtime CPU pressure

[more-shubham]

Worker count is currently a static flag defaulting to 3. On multi-core machines or varying I/O conditions the optimal count changes during a run. A runtime tuner makes the system self-optimizing.

Scope:

• Create internal/tuner/tuner.go with a background goroutine sampling throughput every 10 seconds
• Track images/sec over a sliding window
• Scale workers up by 1 if throughput improved in last window and memory is below 60% threshold
• Scale workers down by 1 if throughput degraded or memory exceeds 70% threshold
• Respect min/max worker bounds (-min-workers, -max-workers flags)
• Log every scaling decision with reason

Example log output:

[Tuner] Scaled workers 3 → 4 (throughput: 6.08 → 7.21 img/s, memory: 42%)
[Tuner] Scaled workers 4 → 3 (throughput: 7.21 → 6.89 img/s, cause: diminishing returns)

Acceptance Criteria:

• Worker count adjusts automatically without restart
• Scaling decisions are logged with throughput and memory context
• Static -workers flag still works, disabling the tuner
• Benchmark comparison between static and adaptive modes documented in BENCHMARKS.md