Add rotate-and-reduce for linear sequences of rotations

[j2kun]

Note: rebased off #530

So this is working but... I can't say I'm particularly happy with it.

Now it recognizes sequences of rotates-and-adds that comprise an aggregation of every value in a vector (exactly once), and replaces them with the logarithmic rotate-and-reduce trick. I wrote down a few non-triggering examples, but I'm not confident it's bug-free by any means.

To the extent that this PR uses a dataflow analysis, I'm somewhat happy. It's an improvement over how I handled the version of this trick that worked on tensor.extract. However, the dataflow analysis has some limitations that left me frustrated. Specifically, I wanted to have the dataflow analysis accumulate the tensor accesses with multiplicity and across multiple tensors. The original
member variable inside RotationLattice was something like DenseMap<Value, std::unordered_multiset<int64_t>>. The join method then updated all accessed tensors and added the index access counts rather than just taking a union as it does in this PR.

However, doing this hit assertions in the dataflow framework because the Lattice class asserts that repeated applications of join are idempotent (they call it monotonic) but in my case that would accumulate. I worked around it by constructing the analysis to work with set unions, and then added an extra step in the actual pass to do more specific detailed checks for validity. I hoped that with the extra index multiplicity information, we might be able to do more sophisticated transformations, like reordering some of the ops instead of just giving up. E.g., one thing that wouldn't work here is if the IR computes twice the sum of the tensor elements. I wonder if not using the Lattice framework and going one step lower to AbstractSparseForwardDataFlowAnalysis might fix that, or just making a totally custom analysis unrelated to the dataflow engine, but I ran out of steam this week. I also suspect we could rely on some more academic ideas like e-graphs to identify more optimal transformations.

Fixes #520

Fixes #511

[asraa]

Thank you! Just a few nits