Shift --secret-to-<scheme> passes from accepting arith to requiring modarith

[AlexanderViand-Intel]

Inspired by @WoutLegiest 's PR on arith-to-mod-arith:

• Arith to ModArith Conversion pass and Mac Transformer pass #1125

In the past, we've discussed that there's a disconnect between various high-level arithmetic operations (e.g. a + b in Python, arith.addi %a, %b : i32 in high-level IR, etc) and what FHE actually does, especially in the arithmetic schemes, where the plaintext modulus is virtually never a nice power of two. In binary schemes, there's less of a disconnect, but it's still there (e.g. integer-level RNS?).

Since we're now (again, mostly for the arithmetic schemes) starting to more clearly differentiate the "arithmetization"/"binarization" phase from the actual scheme/LWE level lowerings, I think it might be a good idea to make this conversion from application-level semantics to plaintext-level semantics more explicit and only accept "mod arith with modulus that matches plaintext modulus" for the (arithmetic) secret-to-scheme lowerings.

This not only reduces possible errors/footgun situations, but it also means we could let people run the actual "plaintext" circuit to verify that arithmetization (including changing from Int or i32/etc to plaintext modulus) doesn't mess up their application logic.