Support larger matvec/conv that are split across multiple ciphertexts

If the # of elements are too many to fit in a ciphertext, we currently don't write out kernels to support splitting them across many kernels. This would require mostly supporting larger matvecs, since we can reduce convs to matvecs, but it would still require plumbing, testing, and making sure that the layouts expect (ct, slot) with ct > 1.