Add OllamaLiteClient for zero-dependency Ollama access

[vijaytank]

Problem

Users running Ollama from CLI tools, edge utilities, or lightweight apps hit several compounding issues with the full Spring AI Ollama stack:

• The WebClient/RestClient transitive dependency graph is prohibitively heavy for callers that only need a single blocking generate() call (see Use JDK HttpClient with OpenAI SDK instead of OkHttp3 #5838).
• WebClient read timeouts fire immediately on slow/large-prompt Ollama responses with no straightforward configuration path (see When calling the Ollama chat model, Ollama responds slowly, causing a read timeout error #3573).
• Reactive cancellation signals do not cleanly terminate the underlying Ollama inference thread, causing GPU resource leaks (see After an interrupt request, the request to Ollama was not properly terminated. #2351, Ollama inference thread not terminating after HTTP request is interrupted, causing GPU resource leak #2944).

Solution

OllamaLiteClient is a zero-external-dependency alternative that uses only java.net.http.HttpClient (JDK 11+). It provides a single blocking generate(model, prompt) method targeting /api/generate, with proper input validation and safe JSON character escaping.

It lives alongside the existing OllamaApi in the org.springframework.ai.ollama.api package — users who need the full Spring stack (streaming, observability, retry, auto-configuration) should continue using OllamaApi + OllamaChatModel.