Fix FasterCache crash and SmoothedEnergyGuidance hook install bugs

[cursor]

Summary

Fixes two critical correctness bugs found during automated review, plus a small tooling fix for copy checking in environments where ruff is not on PATH.

Bug 1: FasterCache crash on guidance-distilled models

Impact: Using FasterCache on Flux and other guidance-distilled transformers crashes on the first denoising step inside the configured attention skip timestep range.

Root cause: FasterCacheBlockHook attempted to read self.state.cache[-1] before any forward pass populated the cache. For guidance-distilled models, is_guidance_distilled=True forces attention skip even on iteration 0 when batch sizes match.

Trigger: FasterCacheConfig(is_guidance_distilled=True, spatial_attention_timestep_skip_range=(t_low, t_high), ...) with a timestep inside the skip range on the first block forward → TypeError: 'NoneType' object is not subscriptable.

Fix: Only allow attention skip when self.state.cache is not None, so the first forward computes normally and seeds the cache.

Validation: pytest tests/hooks/test_faster_cache.py — 1 passed

Bug 2: SmoothedEnergyGuidance completely broken

Impact: Smoothed Energy Guidance never installs query-blur hooks, so SEG has no effect regardless of configuration.

Root causes (two bugs):

1. SmoothedEnergyGuidance.prepare_models overrides the base method without incrementing _count_prepared, so the hook gating condition is_conditional and _count_prepared > 1 is never satisfied.
2. The seg_guidance_layers shorthand builds SmoothedEnergyGuidanceConfig(layer, fqn="auto"), passing an int to indices (which must be list[int]), causing TypeError during hook install — same class of bug as PR Fix LayerSkipConfig shorthand crash in SkipLayerGuidance and AutoGuidance #10 for LayerSkipConfig.

Fix:

• Add self._count_prepared += 1 in prepare_models.
• Pass indices=[layer] when building configs from shorthand.

Validation: pytest tests/guiders/test_smoothed_energy_guidance.py — 2 passed

Tooling

• utils/check_copies.py: invoke ruff via python -m ruff so pre-commit copy checking works when ruff is not on PATH (same approach as PR Fix LayerSkipConfig shorthand crash in SkipLayerGuidance and AutoGuidance #10).

Existing open PRs (unchanged)

These tracked bugs still have open PRs awaiting review:

• PR Fix variant detection crash and parallel shard loading race #5: variant detection crash + parallel shard loading race
• PR Fix swapped tensor order in TransformerBlockSkipHook for SkipLayerGuidance #7: TransformerBlockSkipHook swapped tensor order
• PR Fix LayerSkipConfig shorthand crash in SkipLayerGuidance and AutoGuidance #10: LayerSkipConfig shorthand int vs list (SkipLayerGuidance/AutoGuidance)