[Transforms] Add constant_tensors_folding pass

[niuxiaog]

This PR implements constant_tensors_folding pass (RFC: PR #183) in issue #56 and issue #146:

• The input MLIR function entry will be split into two functions: entry() and runtime_fold(). The runtime_fold() function contains the constant operations whose input tensors and output tensors are all constant. The new entry() function contains other operations that depend on the variable values and folded constant tensors.
• When needed, compile_time_fold() can be enabled to fold constant tensors during compile time.
• A constant tensor may be fully folded by several sequential operations. If fully folding increases the data size dramatically (i.e., by BroadcastOp), we choose a partial folding.
• Swaps the constant BroadcastOp and the constant operations after it, so that these constant operations can be included into partial folding. Currently there are strong constraints on these ops to ensure correctness.
• Necessary information for runtime execution is added as GlobalOp to the MLIR module.
• During the pass, the buffers for storing folded tensors are allocated using the APIs provided by the constant cache manager, which will be implemented by another PR [Runtime] Constant cache manager and runtime pipeline #342.

[ciyongch]

Is there any shortcut for this kind of operation?

[ciyongch]

Shall we combine these two passes into one, and provide an option to do the analysis only is needed?

[Menooker]

I feel the same. Maybe we should put the analysis into the pass, unless the const-subgraph-analysis is needed by more than one pass.

[niuxiaog]

OK, I will put them into to one.

[niuxiaog]

Though the PR is ready_to_review, we need PR #342 and benchgc's support to fully enable this feature. For OpenVINO integration, this feature can be enabled by modifying file.