Add tensor_ext.combine and an associated pattern to replace collapse-insertion-chains

[j2kun]

From the review on #471

In HECO, "left-over" insert/extract ops instead get mapped to the combine op which lets you merge two tensors. The combine op essentially models the "masked addition" you sometimes get in batched FHE code: given a (plaintext) mask m, you do a * m + b * !m, letting you pick independently for each index whether it comes from a or b. While there's of course a canonicalization to handle the "full replacement" case that this pass handles, not all programs batch that nicely (of course, in the HECO examples, we focused on those that do 😉). Because an insert/extract generally costs one (ptxt-ctxt) mask multiplication and one (much, much more expensive) rotation, mapping things to combine instead gives much better performance for those programs that don't batch perfectly. In addition, the combine op is pretty useful in itself to express batched FHE programs, so it'd be nice to add to tensor_ext. Unfortunately, the HECO implementation of combine has a rather messsy parser/printer and canonicalizer (though it gets the job done).

[github-actions]

This issue has 1 outstanding TODOs:

• lib/Dialect/TensorExt/Transforms/Passes.td:32: Investigate replacing this pattern with a tensor_ext.combine op

This comment was autogenerated by todo-backlinks

[j2kun]

Part of #475