Normalize decoded Edwards25519 coordinates before arithmetic

[ladnir]

A decoded point could round-trip correctly through the assembly backend yet produce an incorrect cofactor multiple. This caused a deterministic SimplestOT mismatch at index 7,200 in a 38,448-OT batch and subsequently broke a RingLPN correlation.

Normalize x with one four-lane field multiplication after decoding's sign selection, before subsequent point arithmetic. Add a regression for the failing point with both signs in all eight lanes, comparing cofactor clearing against the scalar implementation.

Validation:

• The added regression fails before the fix and passes afterward.
• GCC 13/ASan: 33 focused curve, base-OT, RingLPN, RevCuckoo, and AnyField tests passed; 3 configuration skips.
• The standalone 38,448-OT SimplestOT reproducer now reports zero mismatches.
• Windows: both Edwards25519 test groups and RingLPN base generation passed.

Consumed by osu-crypto/libOTe#188 and its stacked PRs #189 and #190.