Stop tracking model weights, checkpoints, and training data

Weights and optimizer state are multi-GB, regenerate from the training
scripts, and stay in history forever once committed -- .git is 21GB, of
which three Qwen3 CPT checkpoints alone are 10GB.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017hfvdqpbU5UVf8PmzbvZms

.gitignore

@@ -1,2 +1,13 @@
 .vscode/
 __pycache__/
+
+# Model weights and optimizer state: multi-GB, reproducible from the training
+# scripts, and they permanently bloat history once committed.
+*.safetensors
+*.pt
+*.pth
+*.bin
+*.arrow
+*.parquet
+checkpoint-*/
+llm/data/