ModelMeta: expose chat template, special tokens, and full GGUF metadata

[vaiju1981]

Summary

Enriches the existing getModelMetaJson native method and the ModelMeta wrapper with three
read-only introspection fields, so callers can inspect a loaded model without a second round-trip:

• chat_template — the model's resolved Jinja template (ModelMeta.getChatTemplate())
• special_tokens — bos/eos/eot (plus sep/nl/pad in the JSON) via getBosTokenId() /
  getEosTokenId() / getEotTokenId()
• metadata — the full GGUF key/value map via getMetadata(key) (e.g. "general.architecture"),
  capped at 2 KB/value so large array metadata (tokenizer tokens/merges) can't bloat the payload

Why

ModelMeta already exposes n_ctx_train / n_embd / n_vocab / n_params / size /
architecture / name. These three fill the common remaining gaps: callers frequently also need
the model's own chat template (to drive prompting), its special-token IDs (chat/FIM), and arbitrary
GGUF metadata — all of which previously required going outside the binding.

Design

• No new native method or JNI signature change — the fields ride the existing getModelMetaJson
  JSON payload, and the new ModelMeta getters default cleanly when a field is absent (so a mismatched
  older native lib still works).
• The special-token map is shared between the normal and vocab-only paths via a small
  special_tokens_json helper.
• The 2 KB/value metadata cap is deliberate — scalar metadata fits; huge array values truncate rather
  than dumping the whole tokenizer into the JSON.

Testing

• ModelMeta unit tests (JSON-based, no model): 11 total, green — 2 new cover the added getters
  and their absent-field defaults.
• Native: libjllama builds and links against b9842 and loads cleanly (NativeLibraryLoadSmokeTest).
  Field population over a real GGUF is exercised by CI's model-backed suite.

[bernardladenthin]

Hey again @vaiju1981! Great PRs. I'll move now many files to integrate the longchain as artifact. May wait some hours with new code changes. It will be hard to merge :)