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

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`.

Start by tracing the lowering of LAPACK routines to the generated LLVM function wrappers and the jit_call path. Replace hardcoded matrix sizes with size arguments passed to jit_call, and verify that dynamic-size matrices no longer require one wrapper per instance.

Written by the indexing model from the issue text.