Improve `LLamaEmbedder`

[martindevans]

Description

The LLamaEmbedder at the moment does not expose all of the embedding capabilities of llama.cpp.

Currently:

• It only returns 1 single vector as a float[]
• Some models return a single vector which represents the entire input sequence (embeddings models) and some produce an embedding vector per token (generative models).
• Pooling mode can be set with some models which sets a method for converting lots of embeddings into one embedding. Probably only compatible with some models?

Improvements:

• Indicate which type of model an embedder was created with.
• Indicate how many results there are.
• Use llama_get_embeddings, llama_get_embeddings_ith and llama_get_embeddings_seq as appropriate to get the correct embeddings.

Random things to consider in no particular order:

• What should be returned? float[][] or a Span<float>?
• When tokens are added to the batch, should the logits flag be set for all tokens, none of the tokens, or just the last token?
  • Does this change with different models (generative vs embedding vs generative with pooling)?
• Should returned embeddings always be normalized?

[martindevans]

@Lyrcaxis Some notes on LLamaEmbedder improvements, as discussed in Discord.

[martindevans]

See #902 for a start on some of this.