Currently the double_keccak() ARM intrinsic uses an implementation of SHA3 that relies on both Neon and specialized cryptographic instructions that are not available on many platforms, including Neoverse. This causes ML-KEM to use the Java implementation of double_keccak() on these machines, taking a large performance hit as a result. To fix this, a second implementation is required that only uses GPRs and the generate_sha3_implCompress_gpr() intrinsic.