Unknown bug related to value capture

[AntonLydike]

When running against upstream LLVM/MLIR tests, I see a bunch of very similar failures, which I suspect are due to the same bug. I am not quite sure what's wrong yet though.

Currently responsible for 9 / 69 remaining test failures:

• mlir/test/Dialect/Bufferization/canonicalize.mlir:362
• mlir/test/Dialect/Bufferization/Transforms/buffer-deallocation-simplification.mlir:51
• mlir/test/Target/Cpp/for.mlir:68: error: Couldn't match "int32_t [[SN:[^ ]*]] = add([[SI]], [[ITER]]);". (here SI was captured empty for some reason?)
• mlir/test/Dialect/Affine/slicing-utils.mlir:65
• mlir/test/Dialect/LLVMIR/roundtrip.mlir:619
• mlir/test/Dialect/Bufferization/Transforms/OwnershipBasedBufferDeallocation/dealloc-region-branchop-interface.mlir:55
• mlir/test/Dialect/Bufferization/Transforms/OwnershipBasedBufferDeallocation/dealloc-existing-deallocs.mlir:22
• mlir/test/Dialect/Bufferization/Transforms/OwnershipBasedBufferDeallocation/dealloc-callop-interface.mlir:31
• mlir/test/Dialect/Bufferization/Transforms/OwnershipBasedBufferDeallocation/dealloc-branchop-interface.mlir:42