fix: make base-model warning's instruct advice family-aware

[inureyes]

Summary

The no-chat-template warning that mlxcel run / the chat REPL prints when a model ships no chat template advised users to "try a variant named with an -it suffix". That is the Gemma naming convention only. For other families the advice was actively misleading: Llama and Qwen2.5 instruction-tuned checkpoints use -Instruct, and Qwen3 / Qwen3.5 use the plain repo name with -Base marking the non-instruct variant. So a user running Qwen3.5-0.8B-Base (the base sibling of this repo's own README test model Qwen3.5-0.8B) was pointed at a non-existent -it repo instead of simply being told to drop -Base.

The advice now states the per-family conventions:

• Gemma → -it suffix (e.g. gemma-4-e4b-it-4bit)
• Llama / Qwen2.5 → -Instruct
• Qwen3 / Qwen3.5 → plain name, with -Base marking the non-instruct variant

Base-model detection is unchanged — it keys on chat-template absence (processor.is_none()), never on the model name. Only the human-facing advice text changed.

Changes

• src/commands/chat.rs — replace the -it-only advice block in the base-model warning with family-aware guidance.
• CHANGELOG.md — [Unreleased] → Fixed entry.

Tests

• cargo fmt --check — clean.
• cargo check --features metal,accelerate --bin mlxcel — compiles. No test asserts the warning wording, so no test change is required.

Related

Follow-up to the base-model warning introduced in #134 and the structured User/Assistant fallback in #136.