Evaluate OV+GC integration on CPU

[ZhennanQin]

1. Build OV+GC from https://github.com/slyalin/openvino/tree/mlir.
2. Collect MLP performance with https://github.com/slyalin/openvino/tree/mlir/tools/mlir_bench, including OV-baseline, OV+mlir-tpp, OV+GC.

[AndreyPavlenko]

1. Build OV+GC from https://github.com/slyalin/openvino/tree/mlir.

It seems to be done - slyalin/openvino#155

[niuxiaog]

It seems to be done - slyalin/openvino#155

There are some issues with the current integration:

1. The dlti information is not attached to the mlir module.
2. It can not build successfully with current main branch and zhicong/perf_test branch. Link errors exist.
3. MLIR-based pipelines, including tpp-mlir and gc-mlir, can not execute with multi-threads. The -nthreads takes effects only on OV native execution. OMP_NUM_THREADS has no effects.

[ZhennanQin]

MLIR-based pipelines, including tpp-mlir and gc-mlir, can not execute with multi-threads. The -nthreads takes effects only on OV native execution. OMP_NUM_THREADS has no effects.

Please rebuild OV with -fopenmp flag. You need to do this by hacking OV's cmake file.

[niuxiaog]

Please rebuild OV with -fopenmp flag.

-DTHREADING=OMP works. And modification on tpp is also needed.

[niuxiaog]

Here is a usable integration: slyalin/openvino#167. Compared with slyalin/openvino#155, these are the updates:

1. DLTI information is attached to llvm module.
2. Support for constant tensor folding is added, including marking constant input tensors, parsing relevant information from llvm module optimized by GC, allocating buffers for folded tensors and calling fold and entry funcs correctly. This is an experimental implementation and will be optimized later.
3. Scripts for benchmarking is updated to support multi threading.
4. Some experimental codes for testing, such as generating constant weights for mlp, splitting out the TransposeOp from linalg::MatmulTransposeBOp.
5. Building and linking errors are fixed.

TODO:

1. Accuracy test.

[lmontigny]

pipeline OV + graph v2 = ready for work.
constant weight -> to negotiate with OV