Perf: load weights, create KV cache, initialize tokenizer in parallel

[Bekaboo]

Use multiple thread to load weights, cache and tokenizer, should slightly improve the initialization and TTFT time.