Cohere Command-R runner.
Command-R / Cohere2 ship as general.architecture = command-r or
cohere2 in their GGUF converters — Llama-shaped with
parallel-residual attention and no embedding-output norm. This crate
is a thin wrapper over [rlx_llama32::Llama32Runner] with arch
validation.
Caveat: Command-R's parallel residual (Q/K/V and FFN added in
one residual pass) and missing embedding-output LayerNorm aren't
yet wired in rlx-llama32 — runs will produce some tokens but
won't match the upstream reference until those land. PLAN.md M4
follow-up.