Most implementations of md5 internally consist of 3 functions: md5_init - which initializes internal context, md5_update - which accepts portions of data and processes them and md5_final - which finalizes the hash and releases the context. These roughly suit aggregate's internal functions (SFUNC and FINALFUNC, md5_init is probably to be called on first actual input). Since performance is important for you the functions should be written in low-level language as C, to me it doesn't look difficult to take some C md5 module and adapt it to be an aggregate... though it's not like I would do this easily myself :)
