[Target] Implement pass inserting tensor.pad if required

In the future we will have scenarios requiring padding to be inserted to be able to effectively tile and vectorize operations.

This pass implements the logic required to insert and optimize `tensor.pad` operations to pad operands of linalg operations to a multiple of a given number. The current logic only does so for tile sizes.

Optimization wise, the pass optimizes zero-padding to undef-padding where possible and attempts to fuse them into various operations when possible.

A future pass will lower `tensor.pad` to DMA transfers to fully integrate it into compilation flow.

codegen/compiler/src/Quidditch/Target/CMakeLists.txt

@@ -25,6 +25,7 @@ iree_cc_library(
         "ConfigureForSnitch.cpp"
         "DisableQuidditchVariant.cpp"
         "LinkExecutables.cpp"
+        "PadToTilingConfig.cpp"
         "ReluToMax.cpp"
         "RemoveTrivialLoops.cpp"
         "TensorTile.cpp"