[Tracking] PT model support follow up

[masahi]

#207 is only the first cut. Many TODO items are left

• Fix memory profiling Enable running PyTorch models  #207 (comment)
• Make single-gpu performance at parity with the MLC model
• Make multi-gpu performance sane
• Consider using cuda graph if we decide to keep the 2D padded input representation
• Or, consider reverting the 2D input change
• Revisit custom changes to our vllm fork https://github.com/octoml/vllm/tree/for-mlc-serve and minimize them
• Figure out how to support other models besides the ones in vllm
• Support parallel-sampling eviction by recompute (requires model change)