Fix SEG _gaussian_blur_2d: finite blur_sigma is ignored (inverted infinite-blur branch)

[XIONGPEILIN]

What does this PR do?

Fixes a branch inversion in SmoothedEnergyGuidanceHook / _gaussian_blur_2d
(src/diffusers/hooks/smoothed_energy_guidance_utils.py) where the finite-sigma and
infinite-blur code paths are swapped. As a result any finite blur_sigma is silently
discarded and replaced by a uniform (spatial-mean) query, so sweeping finite sigma has no effect.

The bug

is_inf = sigma > sigma_threshold_inf
...
if is_inf:                       # sigma "infinite"
    ... 2D Gaussian conv ...     # <- should be the uniform/mean case
else:                            # finite sigma
    query_slice[:] = query_slice.mean(dim=(-2, -1), keepdim=True)   # <- discards sigma

This is inverted relative to:

1. This code's own documented behavior. The guider/pipeline doc for seg_blur_sigma states:
   "Setting this value greater than 9999.0 results in infinite blur, which means uniform queries."
   i.e. sigma > threshold should give the spatial mean, and a finite sigma should be an actual
   Gaussian blur.
2. The original SEG-SDXL implementation this function is "Copied/Modified from"
   (pipeline_seg.py#L171-L175):

   if not self.inf_blur:   # finite sigma
       query_ptb = gaussian_blur_2d(query_ptb, kernel_size, self.blur_sigma)
   else:                   # infinite
       query_ptb[:] = query_ptb.mean(dim=(-2, -1), keepdim=True)

Because the default seg_blur_sigma=9999999.0 always takes the (huge-sigma ≈ uniform) conv branch,
the bug is masked for default usage and only surfaces when a finite sigma is requested.

Reproduction

import math, torch
from diffusers.hooks.smoothed_energy_guidance_utils import _gaussian_blur_2d

q = torch.randn(2, 1024, 8)  # 1024 = 32x32 image-token grid
def run(sig):
    ks = math.ceil(6 * sig) + 1 - math.ceil(6 * sig) % 2
    return _gaussian_blur_2d(q.clone(), ks, sig, 9999.9)

print(torch.equal(run(4.0), run(16.0)))   # main: True  -> finite sigma has no effect

On main, different finite sigmas are byte-identical (max|sig4 - sig32| = 0.0). With this PR they
differ (larger sigma smooths more), while sigma > threshold returns the exact spatial mean.

Fix

Swap the two branches: is_inf → spatial mean (exact uniform queries); finite sigma → the real 2D
Gaussian blur of standard deviation sigma.

Note on behavior change

For the default (sigma > threshold) the output changes from a huge-kernel Gaussian conv (an
approximate uniform) to the exact spatial mean — matching the documented "uniform queries".
Finite blur_sigma values now actually control the blur amount. SEG is marked experimental in the
code, so this may affect users who relied on the (buggy) finite-sigma-as-uniform behavior.

Before submitting

• Read the contributor guideline.
• ruff check / ruff format pass on the changed file.
• Happy to add a small regression test (finite sigmas differ, inf == spatial mean) if preferred.