Update openvino-mlir-gc integration

[niuxiaog]

No description provided.

[dchigarev]

@niuxiaog could you please briefly describe what this PR does?

[niuxiaog]

@niuxiaog could you please briefly describe what this PR does?

Sure. I will update the RFC, which introduces the constant tensor folding pass and the related parts in OV integration.

[niuxiaog]

@niuxiaog could you please briefly describe what this PR does?

1. DLTI information is attached to llvm module. Currently, default values are from SPR.
2. Support for constant tensors folding pass 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. Building and linking errors are fixed to integrate latest main branch of GC.
5. Some experimental codes for testing, such as generating constant weights for mlp, splitting out the TransposeOp from linalg::MatmulTransposeBOp. These codes will be deleted before merging.

For a more complete introduction of constant tensors folding pass, please refer to RFC.

[dchigarev]

why do we need to split it?

[niuxiaog]

This is experimental code to test that the TransposeOp can be moved to fold() by the constant tensor folding pass. It will be deleted later.