I rebase the branch gniibe/t8204 to master.
The commit rC320f0d209bff: Clean up unused functions. has compile-time selection with AVX2.
Then, I realized that libgcrypt always offers runtime selection of such CPU specific feature.
Thus, I modified the implementation like:

• optimized version should have variant for these three functions: kyber_keypair, kyber_encap, kyber_decap
• In kem.c, checking the hardware feature, select relevant variant for that

In the reference implementation of Dilithium, AVX2 version has a commit of 5eaf6e94129967a89f7f2cb51006ff2666beac4d, for better data locality in signature generation.

To respect this improvement, we don't share crypto_sign_signature_internal and crypto_sign_verify_internal between non-vector version and AVX2 version.