perf(stark): optimize non-packed constraint evaluation

Packing during constraint evaluation requires the prover to allocate temporary buffers for storing the transposed field elements. Creating these requires a linear pass over the traces before even starting constraint evaluation. In the scalar path, we can just pass a reference to the row, which would reduce allocation and remove a linear pass over the trace (both being particularly useful for WASM).

---

Ported from https://github.com/0xMiden/p3-miden/issues/51