support dynamic-size matrices on lowering of `lapack` routines

[mofeing]

currently, we generate one LAPACK LLVM function wrapper per instance because we hardcode the sizes of the matrices on it. this could be better handled by just passing the sizes as arguments to the jit_call.