Introduce MLIR transform dialect to BladeDISC

[wyzero]

We'll start to explore using MLIR transform dialect to do codegen for (fused) compute-intensive pattern. The initial target is to support gemm codegen on ARM platform to address the dynamic shape problem of Arm Compute Library.

The initial plan is:

• Step 1, enhance the fusion decision pass. We’ll add a new fusion kind kTransform for the transform-based fusion pattern.
• Step 2, lower the lmhlo fusion op to linalg on tensor.
• Step 3, transform the linalg computation to loops using transform dialect.
• Step 4, refined the transformed loop to make it suitable for BladeDISC runtime.
• Step 5, add a new pass to the disc pass pipeline to drive the above process.
• Step 6, weight pre-packing support
  • add disc_linalg.multi_level_pack op, used for doing packing.
  • add transform.disc.cache_read transform op, relying on disc_linalg.multi_level_pack op.
  • add folding support for disc_linalg.multi_level_pack.
  • lower disc_linalg.multi_level_pack to loop if it can not be folded.
  • fuse const weight op into the kTransform fusion pattern, lower it to linalg and then schedule it.
• Step 7, assign a default schedule for each kTransform pattern.
• Step 8, schedule selection logic injection
• Step 9, initial model level testing: bert (albert).
• Step 10, support nt, tn, tt format GEMM.
• Step 11, support batch matmul
• Step 12, support GEMM epilogue fusion.
• Step 13, performance optimization