Add DP supervised fine-tuning module for Gemma models

[copybara-service]

Add DP supervised fine-tuning module for Gemma models

Introduces dpsynth.text.model and dpsynth.text.dp_sft, a composable API for
differentially private supervised fine-tuning of Gemma language models with
LoRA adapters. Split into two files to separate concerns:

model.py (no DP logic, standard ML review):

• SupportedModel: Enum of validated Gemma variants.
• LoraConfig: LoRA adapter configuration.
• load_gemma(): Load pretrained model with LoRA adapters applied.
• sft_loss_fn(): Cross-entropy loss (forked from tunix peft_trainer).

dp_sft.py (DP-critical, needs privacy review):

• DPSft: DPMechanism subclass wrapping DP-SGD via JAX Privacy.
• calibrate(): Noise calibration from a zCDP budget.

dp_sft_test.py:

• Unit tests for loss function, calibration, config, and enum wiring.
  All tests use a tiny mock model (no real checkpoint loading).

Composes Tunix (model), qwix (LoRA), and JAX Privacy (DP training).