[trainer] fix: update TorchTitanEngine for latest torchtitan API

[acisseJZhong]

Summary

• Update TorchTitanEngine to align with torchtitan HEAD API changes: remove deleted expert_tensor_parallel_degree/etp/maybe_enable_amp, add concrete CrossEntropyLoss.Config for new abstract BaseLoss
• Fix attn_type being silently ignored ([trainer] bug: TorchtitanEngine silently ignores attn_type="flex" — no clear BKM for which torchtitan version to use #6182): pass attn_backend= to model_registry() and fix wrong attribute path in prepare_model_inputs
• Fix derive_torchtitan_name_and_flavor to handle config factories (callables) and len(layers) fallback for layer count matching
• Fix test script: use NUM_GPUS for rollout TP size instead of hardcoded 8, fix misleading experiment name

Fixes #6182

Test plan

• Verify derive_torchtitan_name_and_flavor correctly resolves Qwen3-0.6B flavor
• Run tests/special_e2e/run_ppo_trainer_torchtitan.sh with torchtitan HEAD
• Verify model builds with attn_type=flex (FlexAttention) as configured

BACKEND=torchtitan FSDP_SIZE=1 NUM_GPUS=1 MODEL_ID=Qwen/Qwen3-0.6B
   bash tests/special_e2e/sft/run_sft_engine.sh 

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[gemini-code-assist]

Code Review

This pull request updates the torchtitan engine implementation and its associated E2E test script. Key changes include refactoring model specification and attention backend handling, removing expert tensor parallelism configurations, and adding a placeholder loss function for Trainer initialization. The model flavor derivation logic was also enhanced to support callable configurations and different layer attribute names. I have no feedback to provide as there were no review comments to assess.