Add DictaLM Qwen3 model profile

[Tovli]

Summary

• add ADR-027 and an explicit, architecture-aware Qwen3 GGUF path
• introduce the pinned DictaLM bundle/profile, dual stop tokens, prompt policy, native FFI/RN entry point, and CLI/benchmark switches
• require an arm64 HighEnd capability, exact model/tokenizer/template identities, and a detached Ed25519 manifest signature before loading

Why

DictaLM 3.0 is a Qwen3 model; the prior Qwen2-only loader could not load it safely or preserve its tokenizer/template/termination contract.

Validation

• cargo test -p el-core -p el-engine-candle -p el-runtime -p el-ffi -p el-chat -p el-bench
• cargo check --workspace
• cargo fmt --check

Follow-up release gates

Physical-device memory/performance, full-artifact parity, and signed release-bundle evidence remain governed by ADR-027.