Add the ability to benchmark multiple models concurrently.

This is useful for benchmarking multiple LoRA adapters.
- Also fix the latency_throughput_curve.sh to parse non-integer request
  rate properly.
- Also added "errors" to the benchmark results.