Lazy-load CLIP language encoder

[ruziniuuuuu]

Summary

• defer CLIP model and tokenizer construction until a non-null language embedding is requested
• cache the initialized encoder for later calls
• remove unused transformer imports

This keeps non-language dataset imports and training offline and avoids an unnecessary 1.7 GB download.

Closes #300.

Test

pytest -q tests/test_lang_utils.py