ramtorch: percentage-based offload fix for text encoder moving to CPU and back inadvertently causing device mismatch error

[bghira]

This pull request makes targeted improvements to how text encoders and linear layers are managed when using RamTorch, particularly around device placement and partial replacement scenarios. The changes ensure better compatibility and correctness when RamTorch is enabled, especially in mixed-device and percentage-based replacement cases.

Device management and RamTorch integration:

• Improved logic in move_text_encoders (in factory.py) to always skip moving text encoders when RamTorch is enabled, regardless of the target device, to support both partial and full RamTorch setups.

Partial RamTorch replacement handling:

• In replace_linear_layers_with_ramtorch (in ramtorch.py), added logic to track remaining eligible layers when using percentage-based replacement, ensuring only the correct subset is replaced.
• Enhanced handling so that, after percentage-based replacement, any remaining eligible nn.Linear layers (or those matching specific patterns) are moved to the resolved device if they are still on CPU, maintaining device consistency.