feat(ltx2): make run_text_encoder_on_tpu default False and dynamically load torchax

[Perseus14]

This PR isolates the torchax dependency in the LTX2 and LTX2.3 pipelines, making TPU-based text encoding opt-in rather than opt-out.
Changes:

• Lazy Imports: Moved torchax and TorchaxGemma3TextEncoder imports inside the if run_text_encoder_on_tpu: block in ltx2_pipeline.py. This removes the hard dependency on torchax at module import time, allowing users without it to run the pipeline seamlessly.
• Config Defaults Updated: Changed the default value of run_text_encoder_on_tpu from True to False in both the Python getattr fallbacks and the ltx2_video.yml / ltx2_3_video.yml configs.
• LTX-2.3 Config Parity: Explicitly added run_text_encoder_on_tpu: False and enable_dynamic_vae_sharding: True to ltx2_3_video.yml to match the base LTX2 config.

[github-actions]

e2e testgrid: https://8bcf50593faf4ea38060e236169827e5-dot-us-central1.composer.googleusercontent.com/dags/maxdiffusion_tpu_e2e/grid