Make AFL use its own minimization (fixes broken AFL minimization regression after new engine impl)

This is similar to variant task breakage (see https://github.com/google/clusterfuzz/pull/1083). This happens since we try to minimize AFL testcases with libFuzzer and code gets confused whether to use new engine impl or old one for reproducing. We cannot something like variant task fix since we need to keep updating testcase and we should be supporting AFL minimization in the first place.