feat: add FFT trainer worker class

[ShubyM]

This PR splits the existing trainer into the shape we need for full fine-tuning. The old trainer was really a LoRA worker because one process owns a base model and serves multiple jobs by creating and switching adapters. Full fine-tuning has a different ownership model where one worker process owns one trainable model for one job. To make that distinction explicit, this introduces a shared BaseTrainerWorker, keeps LoRA-specific adapter management in LoraTrainingWorker, and adds an FFTTrainingWorker for the single-model full fine-tuning path.

Most of the actual training math is shared between the two modes, so this also moves common forward/backward, padding, logprob, batching, and generation code into the base worker. The loss functions are factored into pure tensor operations in losses.py, which makes them easier to test directly and keeps the worker classes focused on model lifecycle and orchestration.

This PR does not wire full fine-tuning into the API server yet. It just adds the worker split and shared math needed for that follow-up.

[droot]

This is great refactor! Thanks!