Add scripts/reproduce.py + benchmark_latency.py; pin reqs to working transformers/torch range; refresh README with adapter overhead numbers

Initial release: SRT-Adapter v8a (peer-review distribution)