output normalize embedding in '/v1/embeddings'

[redlion0929]

Description

Configure server.cpp to output normalized embeddings with the endpoint set to "v1/embeddings".
This PR is related with issue #5954

Test Method

I use below code to check normalized results.

def check_normalize(embedding):
    s = 0
    for i in embedding:
        s += i * i
    print(f"sum: {s}")


response = client.embeddings.create(input = input_texts, model="test")

print(response)

check_normalize(response.data[0].embedding)

Before

CreateEmbeddingResponse(data=[Embedding(embedding=[0.779549777507782, -1.7770930528640747, -0.5943143963813782, ... ,], index=0, object='embedding')], model='test', object='list', usage=Usage(prompt_tokens=0, total_tokens=0))

sum: 85301.80714120051

After

CreateEmbeddingResponse(data=[Embedding(embedding=[0.0026690999511629343, -0.006084587424993515, -0.0020348725374788046, ..., ], index=0, object='embedding')], model='test', object='list', usage=Usage(prompt_tokens=0, total_tokens=0))

sum: 1.0000004424428977

[ngxson]

Quick note: this is maybe a breaking change. Can we add an option to skip normalize, something like --embeddings-skip-norm ?
By default, openai normalize the vector so I prefer to have normalization enable by default in llama.cpp

[ggerganov]

The flag should better be a request option instead of CLI option

[ggerganov]

I avoided an extra copy of the data and moved the normalization into common so it can be reused

We can add the request option for disabling the normalization in a separate PR