fix: respect gradient checkpointing configuration

[XieWeikai]

Summary

• Stop forcing gradient checkpointing on during training.
• Respect the state set by PI0Pytorch.gradient_checkpointing_enable() and gradient_checkpointing_disable().
• Preserve checkpointed execution when gradient checkpointing is explicitly enabled.

Problem

PaliGemmaWithExpertModel.forward unconditionally set gemma_expert.model.gradient_checkpointing to true whenever the model was in training mode. This overrode callers that explicitly disabled gradient checkpointing, including RLinf SFT configurations with gradient_checkpointing: false.

The existing use_gradient_checkpointing expression already reads the expert model flag and training state, so removing the forced mutation lets the configured setting control the forward path.

Validation

• python -m py_compile src/openpi/models_pytorch/gemma_pytorch.py
• git diff --check
• ruff format --check src/openpi/models_pytorch/gemma_pytorch.py
• ruff check matches origin/main: the only diagnostic is the pre-existing FBT001 warning for use_cache on line 97; this patch adds no lint diagnostics.
• Ran RLinf PI0.5 LIBERO SFT with FSDP on 8 GPUs for 20 optimizer steps, global batch size 200, and gradient_checkpointing: false. All ranks reported gradient checkpointing disabled, no forced-enable message appeared, and training completed successfully.