server: Enable mtmd in llama-server /completion endpoint

[92MING]

ref: #13872

Currently passing media(image/audio) to mtmd is only supported under chat/completion in llama-server.
It is still necessary for allowing mtmd in /completion endpoint, since /completion can have more freedom modification in prompt template(e.g. prefix), or even no template for long article completion.

This PR added an extra field medias(list[object]) under /completion, and each media should contains 2 fields:

• type: audio/image(image_url as alias), other unknown types will be ignored
• data: url/base64 of the image/audio.

Then, make sure you have the same number of <__media__> tag in your prompt to label the position.
Here is an example:

data = {
        'stream': False,
        'top_p': 0.95,
        'temperature': 0.8,
        'top_k': 40,
        'prompt': '<start_of_turn>user\nAnalyze the image and provide a short description\n<__media__><end_of_turn>\n<start_of_turn>model\n',
        'medias': [
            {
                'type': 'image',
                'data': <your img_b64 or url>,
            }
        ],
    }
response = requests.post('http://localhost:8080/completion', json=data)

[ngxson]

Instead of duplicating this whole code block, extract it to a general function and reuse it in /chat/completion and /completion. DRY code principle

[ngxson]

Close and replace by #15108