Limit number of sequences

[yelite]

This PR:

• Adds max_num_seq and max_num_seq_per_request to limit number of sequences in the batch and in each request.
• Use max_num_seq to guide the memory profiling to be less conservative, leaving more room for KV cache
• Expose gpu_memory_utilization in engine config

@sunggg @masahi

[masahi]

This should be sent to https://github.com/octoml/mlc-serve/ as well.