[Feature Request] Allow overriding the embeddings encoding_format for OpenAI-compatible LLM providers

Description The openai Python client (used internally by llama-index's OpenAIEmbedding, which paperless-ngx uses for the LLM/RAG index) automatically requests embeddings with encoding_format="base64" whenever numpy is available in the environment. Genuine OpenAI endpoints handle encoding_format="base64" correctly. However, several "OpenAI-compatible" providers do not fully implement it and return a server error instead. For example, with the IONOS AI Model Hub embeddings endpoint, a base64 request fails with: I confirmed that the same request with encoding_format="float" sent directly to the same endpoint works correctly. Since paperless-ngx doesn't expose a setting for this, the LLM index (llmindex_index task) cannot be used at all with providers that don't support base64-encoded embeddings, even though the provider itself works fine for both chat completions and embeddings under encoding_format="float&quot…