Use AES_* when no AES engine is set; use EVP_* otherwise
Benchmarking shows about 5% overhead for using the EVP_* aes functions over using the AES_* functions, when no accelerators are present. This holds true both on openssl 0.9.7 and 1.0.0, and on OSX 32-bit, atom 64-bit, and Core2 quad 8200 systems, so it's probably a fairly consistent finding.
With #4442 (moved), we made our aes code use EVP always, which is a big win with accelerated AES, but a small loss without. Let's fix that.