Implement log probabilities for chat completions

[atbe]

Log probabilities support was implemented for Ollama's OpenAI-compatible chat completion endpoints.

Key changes include:

• Request parameters logprobs (boolean) and top_logprobs (integer) were added to openai/openai.go's ChatCompletionRequest.
• New LogProb and TopLogProb structs were introduced in api/types.go and openai/openai.go to define the log probability response schema.
• llm/server.go's CompletionRequest was updated to include LogProbs and TopLogProbs, and the completion struct now includes CompletionProbabilities for parsing llama.cpp responses. The Completion function now passes n_probs to the underlying llama.cpp server.
• openai/openai.go's ChatMiddleware was modified to extract log probability settings from incoming requests and store them in the Gin context.
• server/routes.go's ChatHandler now retrieves these settings from the context and passes them to the llm.CompletionRequest.
• Conversion logic was added in openai/openai.go (toChatCompletion, toChunk) and server/routes.go to transform llama.cpp's completion_probabilities into the OpenAI-compatible logprobs format, including token, logprob, bytes, and top logprobs.
• Both streaming and non-streaming responses now include log probabilities.
• Test cases were added to openai/openai_test.go, and a LOGPROBS_IMPLEMENTATION.md file was created for documentation.