Fix 2 embeddings-related issues in server.cpp

[k8si]

Misc embeddings fixes:

• Fix Llama.cpp server embeddings always return a 0 filled vector. #303 - /embeddings server endpoint returns actual embeddings rather than 0-vector
• Fix Support BERT architecture in llamafile #322 - Pull in upstream change from llama.cpp to improve locale handling during lowercase. (This was what was causing an error when using a BERT model with the /embeddings endpoint). Now efficient sentence embedding models like all-MiniLM-L6-v2 should work without a problem.

TODO:

• Still need to test calls to the /embeddings endpoint with a batch of texts, thus far I have only tested with single texts (will finish this on Monday)

[jart]

This fix looks great. I've tested and confirmed it works locally using your script. I think this is worth merging. Please send me the tests on Monday. Then we'll close out the issue.