atoma-vllm: Test deterministic output under PagedAttention cuda kernels

[jorgeantonio21]

This can be tested as:

• Run a vllm instance (by cloning the original vllm repo);
• Send multiple requests, in different orders, to different nodes with same GPU models and check for hashes of requests
• Allow for continuous batching, while running tests