Rethinking Few-Shot Image Classification: A Good Embedding is All You Need?