Clean up SwiGLU reference code weight generation

[andrej]

There's an ugly hack in the SwiGLU reference here:

IRON/operators/swiglu_decode/reference.py

Line 30 in 130b6ea

w_gate = torch.randn(N, K, dtype=torch.bfloat16).T * val_range # gate projection

The first two weight matrices are generated with the wrong dimensions, then transposed. The reason I did it this way initially was to ensure identical outputs between the old CMake-based implementation and the new state (generating random weights in the same order to ensure identical inputs). Now that this is verified and the old CMake-based implementation is gone (#37), it's time to remove this hack.

Thanks @asyms for pointing this out.