llama_functions.py backend tasks

[sklam]

Tasks for llama_functions.py that targets wip/llm_intgrate branch:

• make llama_functions.py test-able by pytest ./llama_functions.py. We will need independent verification of the backend features.
• create "inlined" version for the gen_* functions. see patch_inlined_matmul.patch for reference. Convert the existing functions to call the inlined version to populate the functions.
• check matmul implementation.

  StringAttr.get("parallel") for _ in range(dims - 1)  # batch + M + N
                                            ^^^^^^^^ --- should this be `dims` (no minus 1)?
  

• setup benchmark for attention() then play around with pass fuzzing. What passes can get more performance?

Future looking tasks :

• can we use linalg.matmul or even linalg.batch_matmul? The batch version can do broadcast and transpose!
• linalg dialect supports tensor. can we use tensor instead of memref? Need to figure out bufferization. What is the performance of memref based vs tensor based.