[stable-diffusion] Match diffusers Transformer2D GroupNorm epsilon

[saichowdary007]

Summary

Fixes #1434.

Transformer2D is shared by the Stable Diffusion UNet, but its input
GroupNorm relied on MLX's default eps=1e-5. Diffusers' reference
Transformer2DModel uses eps=1e-6, and the mismatch accumulates through the
UNet. Pass the reference epsilon explicitly while leaving the UNet ResNet and
LayerNorm defaults unchanged.

Testing

• cd stable_diffusion && PYTHONPATH=. python -m unittest test_unet -v
• pre-commit run --files stable_diffusion/stable_diffusion/unet.py stable_diffusion/test_unet.py

The regression test constructs a Transformer2D without downloading model
weights and asserts that its GroupNorm uses 1e-6.

Implementation and test preparation used AI assistance; the submitting
contributor reviewed the changed lines and owns the behavior validation.