Add langchain4j-jllama module: in-process LangChain4j adapters

[vaiju1981]

Summary

Adds a new sibling Maven module langchain4j-jllama that adapts a java-llama.cpp LlamaModel to
LangChain4j's model interfaces in-process over JNI — no HTTP hop, no separate llama-server:

Adapter	LangChain4j interface	Backing call
JllamaChatModel	ChatModel	LlamaModel.chat(...)
JllamaStreamingChatModel	StreamingChatModel	LlamaModel.generateChat(...) (token streaming)
JllamaEmbeddingModel	EmbeddingModel	LlamaModel.embed(...)
JllamaScoringModel	ScoringModel (re-rank)	LlamaModel.handleRerank(...)

Design decisions

• Separate artifact, one-way dependency. The module depends on langchain4j-core:1.17.1, but the
  core net.ladenthin:llama binding gains no langchain4j dependency — plain java-llama.cpp users
  never pull langchain4j (or its Java 17 floor) transitively.
• Sibling module, not in the root reactor. The native build/release pipeline is left completely
  untouched; the module builds independently against the published core jar. Targets Java 17
  (langchain4j 1.x baseline; the core stays Java 8).
• Borrowed model lifecycle. Every adapter wraps a caller-owned LlamaModel and never loads or
  closes it. One model can back several adapters.
• For users who already run the OpenAI-compatible OpenAiCompatServer, langchain4j's
  langchain4j-open-ai client already works over HTTP with zero code. This module is for the
  in-process path (desktop / Android / embedded, no socket).

Testing

• Pure message/parameter/response transforms are unit-tested model-free (LangChain4jMappingTest,
  7 tests): role mapping, multimodal→text flattening, sampling-parameter pass-through, finish-reason
  mapping, and rerank index alignment.
• JllamaChatModelIntegrationTest runs a real chat + streaming round-trip and self-skips unless
  -Dnet.ladenthin.llama.model.path=... points at a GGUF (mirrors the existing model-gated tests).
• mvn test: 7 run, 2 skipped, green.

Adapter contracts were verified against langchain4j 1.17.1 source: doChat is the correct override
point, the chat response reports the model's real finish reason (stop/length/tool_calls) and
token usage, and streaming reports failures via onError (the framework does not wrap doChat).

Not yet mapped (documented in the module README)

• Tool calling (ToolSpecification ↔ jllama ToolDefinition) — the main follow-up.
• response_format (JSON mode); multimodal user input is flattened to text.

Notes for review

• CI is not wired yet. The module isn't part of the reactor, so no existing job builds it; the
  tests above were run locally. Happy to add a small job (mvn -DskipTests install for the core →
  cd langchain4j-jllama && mvn test) if you'd like it gated in CI.
• Placement. It is deliberately standalone, so if you'd prefer this live as a separate repository
  rather than in-tree, that's a clean move — let me know your preference.
• REUSE: all sources carry SPDX headers; the README is registered in REUSE.toml.

[vaiju1981]

This PR is just for me to integrate Jllama as native model within Langchain4j. This PR might mean a separate project.

[bernardladenthin]

Hey @vaiju1981 ,

give me some minutes to integrate it well, then maybe do additional work. Thanks and bests!