Configurable fine-tuning API (TrainingParameters + Optimizer)

[vaiju1981]

Summary

Builds a configurable fine-tuning API on top of the training proof of concept (#287): a
TrainingParameters builder — serialized as JSON across JNI, the same idiom as ModelParameters /
InferenceParameters — driving LlamaTrainer.finetune(TrainingParameters).

LlamaTrainer.finetune(
    TrainingParameters.builder()
        .modelPath(Path.of("base.gguf"))
        .trainingFile(Path.of("corpus.txt"))     // or .trainingText("...")
        .outputPath(Path.of("tuned.gguf"))
        .epochs(3).learningRate(1e-5f).optimizer(Optimizer.ADAMW)
        .lrMin(1e-6f).decayEpochs(3f).weightDecay(0.01f)
        .valSplit(0.1f).nCtx(512).nGpuLayers(0).nBatch(256).nUbatch(64)
        .build());

This PR includes the POC commit from #287, so it supersedes it — #287 can be closed in favor of
this (or merged first, in which case this reduces to the FineTuner delta).

What it exposes

Every training knob the ggml-opt path supports:

• Corpus from inline text or a file (read natively, so a large corpus doesn't round-trip
  through the JVM heap)
• Optimizer — AdamW / SGD (Optimizer enum → ggml_opt_optimizer_type)
• Learning-rate schedule — lr0, lr_min, decay_epochs, weight_decay
• Validation split, context size, GPU layers, logical/physical batch sizes

train_engine.cpp parses the JSON config, applies the knobs to common_params, and runs the same
llama_opt_init / llama_opt_epoch loop as upstream examples/training/finetune.cpp.

Verified

• Native library builds and links against b9842; finetuneNative is exported; the library loads
  (NativeLibraryLoadSmokeTest).
• TrainingParameters builder→JSON unit tests pass (3), model-free.
• LlamaTrainer + TrainingParameters + Optimizer compile through the strict Error Prone / NullAway
  pipeline.
• The actual training run is exercised by the model-gated LlamaTrainerIntegrationTest, which
  self-skips unless -Dnet.ladenthin.llama.train.model=… is set.

Deferred to a focused follow-up

• Progress callbacks (native → Java per-epoch loss/progress). ggml_opt_epoch_callback has no
  userdata slot, so a Java callback needs a thread-local trampoline; since it can't be exercised
  without an actual training run, it is kept out of this change to get proper attention on its own.
• LoRA-target selective fine-tuning (a custom llama_opt_param_filter rather than _all) — more
  involved; upstream's own finetune.cpp trains all parameters.