perf: optimize safetensors metadata fetching latency by fetching in parallel

[saquibsaifee]

💡 What: Refactored fetch_safetensors_metadata to use concurrent.futures.ThreadPoolExecutor for fetching config.json, tokenizer_config.json, and safetensors metadata in parallel rather than sequentially.
🎯 Why: To reduce overall network I/O latency when fetching metadata about safetensors. The previous implementation executed three blocking downloads in sequence.
📊 Measured Improvement: Running a basic warmup-based test benchmarking TinyLlama/TinyLlama-1.1B-Chat-v1.0, latency dropped from ~0.46s down to ~0.32s, approximately a 30% reduction in fetch time. The latency improvement applies best for models containing tokenizer_config.json and valid safetensors metadata alongside config.json by combining their network IO time.

[saquibsaifee]

Benchmark can be reproduced using:

import time
import argparse
from src.models.safetensors_metadata import fetch_safetensors_metadata

def benchmark(repo_id: str):
    # Warmup
    try:
        fetch_safetensors_metadata(repo_id)
    except Exception:
        pass

    start = time.time()
    res = fetch_safetensors_metadata(repo_id)
    end = time.time()

    print(f"Time taken for {repo_id}: {end - start:.4f} seconds")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Benchmark safetensors metadata fetch.")
    parser.add_argument("--repo_id", default="TinyLlama/TinyLlama-1.1B-Chat-v1.0", help="Hugging Face repo ID")
    args = parser.parse_args()
    benchmark(args.repo_id)

[saquibsaifee]

@eaglei15 this PR is ready to be reviewed.