fix: add missing style_encoder optimizer step in reference-guided training

[Mukller]

Bug Fix: Style encoder not updated during reference-guided generator training

During the training loop, the style encoder optimizer step is called after
latent-guided generator training but is missing from reference-guided training.

This means the style encoder never learns from reference images during the
alternating training procedure, which contradicts the StarGAN v2 training algorithm.

Before (missing style_encoder.step for ref-guided training):

# Latent-guided:
optims.generator.step()
optims.mapping_network.step()
optims.style_encoder.step()    # ← present

# Reference-guided:
optims.generator.step()
# optims.style_encoder.step() ← MISSING

After (fixed):

# Reference-guided:
optims.generator.step()
optims.style_encoder.step()    # ← added

Closes #179

[Mukller]

Code Review: fix missing style_encoder optimizer step in reference-guided training

Summary

Adds the missing optims.style_encoder.step() call in the reference-guided training branch. Without it, the style encoder's parameters would not be updated during gradient descent, causing the style encoder to remain frozen while the generator trains against it — leading to suboptimal style transfer quality and silent incorrect behavior.

Critical Issues

#	File	Line	Issue	Severity
1	core/solver.py	~135	optims.style_encoder.step() was missing after g_loss.backward(), leaving the style encoder's weights frozen during reference-guided training	🔴 Critical

What the Fix Does

# Before (bug) — style encoder never updated
self._reset_grad()
g_loss.backward()
optims.generator.step()
# ← style encoder step missing

# After (correct)
self._reset_grad()
g_loss.backward()
optims.generator.step()
optims.style_encoder.step()  # ← style encoder now correctly updated

What Looks Good

• Single-line addition, no side effects
• Consistent with the latent-guided training path, which correctly updates both generator and mapping network

Verdict

Request Changes → now fixed. This is a silent training bug: no crash, no obvious error, but the style encoder would not learn during reference-guided training, producing degraded results.