[New Project] Program Partitioning

[wmdi]

In the current implementation, given a large input program, Mirage rely on users to manually partition it into multiple subprograms. For example, in the chameleon-7b demo(https://github.com/mirage-project/mirage/blob/main/demo/reference_mugraphs/chameleon-7b.py), users need to manually divide the computation into combination of 3 types of kernels (2 rms linear, 1 attention). We want to automate and optimize this step.