One-shot bufferization for onnx dialect

[tungld]

Do bufferization in MLIR style for onnx dialect, where we provide an one-short bufferization for onnx dialect to bufferize tensors, which has several advantages:

• Once tensors are bufferized, the relationship between dynamic dimensions is more obvious via memref, which is benefical to the lowering of onnx ops
• It is easy to make onnx-to-krnl work well with Linalg dialect, so that we can selectively low some onnx ops to Linalg when needed.

More information about bufferization in MLIR: https://mlir.llvm.org/docs/Bufferization/

[chentong319]

This is definitely a direction we can explore. Need detailed discussion. For example, we could just split the current onnx to krnl conversion into two step: destination-passing style onnx-op (the allocation part of current conversion) and loop-oriented ops.