refactor(evasion): migrate to new attack/defense architecture

[asim29]

Summary

Docstring cleanup and two minor clarifying comments in the evasion attack and defense. Both classes were already structurally correct after the ABC base classes landed in PR 2.

Stacked on: `pr/4.2-membership-inference`

---

What changed and why

`attacks/projected_gradient_descent.py` — `EvasionPGD`

• Trims verbose RST `:class:~` cross-references to plain type names in the docstring.
• Corrects the `attack()` docstring: return description said "accuracy" but the method returns a `DataLoader` of adversarial examples.

`defenses/projected_gradient_descent.py` — `AdversarialTrainingPGD`

• Same RST → plain-type docstring cleanup.
• Adds two inline comments in `train_robust()` distinguishing the clean-batch gradient update (Step 1) from the adversarial-batch update (Step 2). The two-step structure was non-obvious without context.

---

Test plan

• `uv run pre-commit run --all-files` passes
• `uv run pytest tests/integration/test_evasion.py -v -m integration` passes
• `from amulet.evasion.attacks import EvasionPGD` resolves
• `from amulet.evasion.defenses import AdversarialTrainingPGD` resolves

🤖 Generated with Claude Code