Add XtsMode128<TDataCipher, TTweakCipher> with zero-abstraction overhead

[Copilot]

Refactors XtsMode following CtrMode128 pattern to eliminate virtual call overhead and keep tweak data in registers instead of spilling to stack.

Changes

• XtsMode128.cs: Generic implementation using IBlock16Cipher<T> constraints for devirtualization. Uses Vector128<byte> directly for tweaks instead of VectorBuffer16 to avoid register/stack transfers.

• XtsMode128.Avx2.cs: 8-block processing with Vector256<byte> and Gf128MulV256 using carryless multiply.

• XtsMode128.Avx512.cs: 16/32-block processing with Vector512<byte> and Gf128MulV512.

Usage

// Before (interface-based, virtual calls)
new XtsMode(AesCrypto.CreateCore(key1), AesCrypto.CreateCore(key2))

// After (generic, devirtualized + inlined)
new XtsMode128<AesCipher, AesCipher>(AesCipher.Create(key1), AesCipher.Create(key2))

All XTS logic (XOR-encrypt-XOR, ciphertext stealing) preserved exactly. All methods marked [MethodImpl(MethodImplOptions.AggressiveInlining)].

Original prompt

参考 CtrMode128 重构 XtsMode，不要随意更改/优化逻辑；重点考虑零抽象开销，保持展开保证 inline，避免类似现在的实现会导致数据在寄存器和堆栈来回读写；

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.