Optimize blur: per-window background caching

[wwmaxik]

Summary

Fixes #42 by implementing per-window background fingerprinting for blur cache invalidation.

Problem

Previously, blur cache was invalidated on any surface commit (via global blur_scene_generation), causing blur to recompute even when the background behind a specific window hadn't changed. This led to:

• High GPU usage scaling linearly with number of blurred windows
• 90%+ GPU usage reported with multiple blurred terminals on NVIDIA

Solution

Replace global scene generation tracking with per-window background hash:

• Each BlurCache now stores last_background_hash
• Hash computed from elements behind each window + window rect
• Blur only recomputes when that specific window's background changes

Changes

• Add last_background_hash field to BlurCache
• Implement hash_background_elements() to fingerprint background
• Move behind_starts calculation before cache invalidation check
• Replace scene_generation check with background hash comparison

Test Results (NVIDIA GTX 1650)

Before:

• 1 blurred window: ~30-37% GPU
• Multiple blurred windows: scales linearly (90%+ reported)

After:

• 1 blurred window: ~10-38% GPU (varies with animated background)
• 6 blurred windows: ~29-37% GPU (stable, no scaling!)

Key improvement: GPU usage no longer scales with number of blurred windows.

Notes

• Remaining ~30% GPU usage is from animated background shader (60 FPS rendering)
• With static shader: GPU drops to 2-3% idle
• Debug logs added to track invalidation reasons (can be removed if needed)

[malbiruk]

Thanks for working on this! I tested it on my laptop and on idle the per-window background caching optimization doesn't show a measurable improvement: 1 blurred terminal sits at ~2% GPU usage, and 6 blurred terminals at ~7% -- both with and without the change. I suspect the original 30% and 90% GPU from #42 might be NVIDIA-specific rather than a cache invalidation issue or something else I'm missing...

Could you try reproducing the original issue on an unmodified main branch on your hardware?

[malbiruk]

Active usage seems to improve a bit though, so I'll merge with follow-up commit consolidating the original needs_recompute block with the new one.