Adding Warmup

[kimbochen]

• Added warmup with CLI argument --num-warmups
• Added endpoint ready checking: Client tests endpoint and waits up to 10 minutes
• Updated sequence length generation logic:
  • --random_range_ratio matches vLLM: sequence length is uniformly sampled in [seq_len * (1.0 - random_range_ratio), seq_len * (1.0 + random_range_ratio)]
  • Added iterative encode-decode to minimize token mismatch