Cache dense encoder method

[shanbady]

What are the relevant tickets?

Closes https://github.com/mitodl/hq/issues/8429

Description (What does it do?)

This PR makes it so we cache the dense encoder instance which avoids unnecessary calls to litellm endpoints and also alleviates us from having to pass around a dense encoder instance since calling the dense_encoder method will have no performance hit.

How can this be tested?

1. checkout main
2. try instantiating the dense encoder and see that there are calls to either litellm/ollama or openai depending on your local setup:

from vector_search.utils import dense_encoder

encoder = dense_encoder()
[2025-10-20 17:47:16] WARNING 7118 [root] litellm.py:25 - [0c41fc84b062] - Model nomic-embed-text not found in tiktoken. defaulting to None

encoder = dense_encoder()
[2025-10-20 17:47:20] WARNING 7118 [root] litellm.py:25 - [0c41fc84b062] - Model nomic-embed-text not found in tiktoken. defaulting to None

encoder = dense_encoder()
[2025-10-20 17:47:20] WARNING 7118 [root] litellm.py:25 - [0c41fc84b062] - Model nomic-embed-text not found in tiktoken. defaulting to None

3. checkout this branch and repeat the same and note there is only one call to the endpoint