Server: fix server hangs on empty prompt

[ngxson]

This is a proposal to fix #5724 and #5246

It's buggy when we run a slot with no tokens to evaluate, since some parts of code implicitly expect n_tokens to be > 0

This PR fixes issue when using:

• /v1/embeddings with "input": ""
• /embedding with "content": ""
• /completion with "prompt": ""

[phymbert]

Can we hardcode a default embeddings answer as OpenAI ? Do you think it deserve a small test scenario ?

[ngxson]

I don't think we should hard code the result, as it's vector and not a fixed text value (so we may expect different floating point inaccuracy on different hardwares)

But what we can do for the test is:

• Test if embedding works with empty input { "input": "" }, only to see if it outputs a vector or not. We don't care what's inside the vector.
• Try to get embedding for one-space { "input": " " } and two-spaces { "input": " " } then compare the euclidean distance of the too vector, you can hard code to check if vector1 != vector2 and distance(vector1, vector2) < THRESHOLD, where THRESHOLD can be hard coded. THRESHOLD can be a bit bigger than it need to be, to compensate the inaccuracy among hardwares.

[ibehnam]

@ngxson @phymbert @ggerganov

Can we also fix the issue where incorrect grammar crashes the server? Last I checked there was a boolean check at the top of the server code which validated the input args. I think we could move the grammar validation outside of that and just return None or Error if incorrect grammar is passed.