perf(control): cache upscaled cond_hint across sliding context steps

[neverbiasu]

Description

• Cache upscaled cond_hint tensor in ControlNetAdvanced
• Slice sub_idxs directly from cached tensor during sliding context steps
• Add guarded cache invalidation (check reference, shape, device, dtype)
• Explicitly release cached tensor in cleanup_advanced()
• Verify exact tensor parity and run test suite

Fixes the TODO in adv_control/control.py (L66).

Previously, sliding_get_control re-ran common_upscale (tensor interpolation) on every single sampling step whenever sub_idxs was present. For video or AnimateDiff workflows, this meant tens of redundant upscale calls on identical input tensors.

This PR caches the full upscaled tensor in self._cached_cond_hint. During sliding context steps, self.cond_hint is sliced directly via _cached_cond_hint[sub_idxs].

Cache invalidation handles changes to cond_hint_original identity, batch size, spatial latent size, device, or dtype.

Verification

• Tests: python -m unittest discover -s tests (16/16 passed).
• Benchmark: Upscale calls reduced from 40 down to 1 across a 20-step 2-window sampling run.
• Parity: Tensor outputs are bitwise identical (diff = 0.0).