feat: Implement GeometricSLM ( replace legacy lqa task )

[Concode0]

This PR introduces the foundational implementation of a Geometric Small Language Model (SLM) using the Versor framework, replacing the existing LQA task. It demonstrates the viability and potential of geometric language modeling.

• Stable Training: Confirmed stable gradient convergence without dimension-induced memory explosion.
• Metrics: Achieved an Eval PPL of 384.92 and a Token Accuracy of 12.97% on a Cosmopedia subset.

Environment Notes & Known Issues
This initial implementation was developed and verified on an M4 MacBook Air. Please be aware of the following hardware constraints:

• MPS Instability: torch.linalg operations and torch.compile are currently unstable on Apple Silicon (MPS) backends.
• Recommended Setup: We highly recommend running this in eager mode or, preferably, in a native CUDA environment.
• Memory Management: Due to the high memory consumption of geometric operations, you may need to manually tune ATTN_BLOCK_SIZE according to your available VRAM.