Polynomial approximation patterns

[j2kun]

Part of #1217

This PR adds polynomial approximation patterns for all unary math dialect ops that operate on floats, e.g., math.exp. These should all be elementwise mappable, so extent to tensors and similar for use in the secret+tensor_ext_layer.

The pass supports extracting the desired degree/interval of approximation from the op attributes, and otherwise uses a default [-1, 1] degree 5 approximation.

What's missing:

• Smarter error handling (e.g., someone asks for an approximation of an op outside its domain)
• Binary/ternary math ops, when all but one argument is statically known (filed a TODO)
• Patterns that match on higher level math ops, e.g., stablehlo.relu lowers to something like linalg.map { arith.maximumf }

[AlexanderViand-Intel]

Amazing! Just a nit and one real request: could you also add this pass to the --mlir-to-secret-arithmetic pipeline, and add at least one math.<something> op in an e2e test? Thanks!

[j2kun]

could you also add this pass to the --mlir-to-secret-arithmetic pipeline, and add at least one math.<something> op in an e2e test? Thanks!

I actually cannot, because we don't have a lowering for polynomial.eval yet.

[AlexanderViand-Intel]

could you also add this pass to the --mlir-to-secret-arithmetic pipeline, and add at least one math.<something> op in an e2e test? Thanks!

I actually cannot, because we don't have a lowering for polynomial.eval yet.

:( That's a good reason, though xD