fix(ruvllm): detect tied word embeddings for Llama 3.2 in candle backend

[ohdearquant]

Problem

Llama 3.2 1B/3B tie the output projection to the input embeddings — the checkpoint ships no separate lm_head.weight tensor. The candle backend hardcoded tie_word_embeddings: false, so loading a Llama 3.2 checkpoint fails on the missing lm_head.weight.

Fix

Detect the tie from tensor presence instead of hardcoding:

let tie_word_embeddings = !vb.contains_tensor("lm_head.weight");

Models that ship an explicit lm_head.weight are unaffected (detected as untied, same as before); Llama 3.2 1B/3B now load correctly.

Scope

6 lines, candle_backend.rs only, behind the candle feature. No new dependencies.

Verification

cargo check -p ruvllm --features candle passes clean (candle-transformers 0.9.2).

🤖 Generated with Claude Code