Fix diffusion training and sampling issues

[SoheylM]

Summary

I fixed the diffusion issues reported in #64, #65, and #66.

• I fixed the 1D diffusion training path so designs are normalized before being reshaped and passed into GaussianDiffusion1D.
• I wired the 1D auto_norm argument into both training and evaluation so restored runs use the saved normalization behavior.
• I changed the conditional 2D diffusion default from 250 to 1000 timesteps. With the existing linear beta schedule, 250 steps left sqrt(alpha_bar_T) ~= 0.282, while 1000 steps brings it to ~0.00635, which better matches sampling from Gaussian noise.
• I fixed DiffusionSampler.diffusion_step_sample to use fresh Gaussian noise for the posterior variance term instead of reusing the predicted noise, and to avoid adding noise at t=0.
• I added focused regression tests for the 1D normalization path, the 2D schedule sanity check, and the sampler variance-noise behavior.

Validation

I validated this in the engibench Conda environment after running the documented EngiOpt install flow (pip install -e .).

• python -m pip check passed with no broken requirements.
• python -m pytest -q passed (3 passed).
• ruff check passed on the changed diffusion files and tests.
• ruff format --check passed on the changed diffusion files and tests.
• python -m compileall engiopt/diffusion_1d engiopt/diffusion_2d_cond tests passed.
• I ran a 1D smoke check on real EngiBench airfoil data: one forward loss, backward pass, optimizer step, and sample all completed with finite values and nonzero gradients.
• I ran a 2D conditional smoke check on real EngiBench beams2d data: one forward loss, backward pass, optimizer step, terminal schedule check, and sampler step all completed with finite values and nonzero gradients.

Notes

A full-repo ruff check engiopt tests still reports pre-existing import-order issues in unrelated files. I kept this PR scoped to the diffusion fixes and verified the touched files instead.

Fixes #64
Fixes #65
Fixes #66

[njhoffman11]

These seem reasonable, but I don't see how this would fix the ghosting issue you see. I would suggest trying an updated version of huggingface diffusers, to rule out issues there