Implement log probabilities for chat completions

[atbe]

Log probability support was introduced across Ollama's OpenAI-compatible endpoints.

Key changes include:

• In api/types.go:
  • A new LogProbs struct was defined to hold token-level log probabilities, token IDs, and top-N log probability maps.
  • ChatResponse and GenerateResponse structs were extended with an optional *LogProbs field.
  • Options struct gained LogProbsEnabled (bool) and TopLogProbs (int) fields to control logprob generation.
• In llm/server.go:
  • The internal completion struct was updated to parse log probability data from the backend.
  • The Completion function now forwards logprobs and top_logprobs parameters to the core generation request and populates the LogProbs field in CompletionResponse.
• In openai/openai.go:
  • OpenAI API response models (Choice, ChunkChoice, CompleteChunkChoice) were updated to include Logprobs.
  • Request models (ChatCompletionRequest, CompletionRequest) now accept logprobs and top_logprobs parameters, which are then mapped to the internal api.Options.
  • Conversion functions were updated to correctly map internal api.LogProbs to the OpenAI schema.
• In server/routes.go:
  • GenerateHandler and ChatHandler were updated to include the LogProbs field in the final API responses.

This enables clients to request and receive per-token log probability information (with configurable top-N) in both normal and streaming modes, maintaining compatibility with the OpenAI chat completions schema.