Fix WMO wall collision, normal mapping, POM backfill, and M2/WMO rendering performance

- Fix MOPY flag check (0x08 not 0x01) for proper wall collision detection
- Cap MAX_PUSH to PLAYER_RADIUS to prevent gradual clip-through
- Fix WMO doodad quaternion component ordering (X/Y swap)
- Linear normal map strength blend in shader for smooth slider control
- Enable shadow sampling for interior WMO groups (covered outdoor areas)
- Backfill deferred normal/height maps after streaming with descriptor rebind
- M2: prepareRender only iterates animated instances, bone dirty flag
- M2: remove worker thread VMA allocation, skip unready bone instances
- WMO: persistent visibility vectors, sequential culling
- Add FSR EASU/RCAS shaders

include/rendering/m2_renderer.hpp

@@ -122,6 +122,7 @@ struct M2ModelGPU {
     bool isKoboldFlame = false;     // Model name matches kobold+(candle/torch/mine) (precomputed)
     bool isLavaModel = false;       // Model name contains lava/molten/magma (UV scroll fallback)
     bool hasTextureAnimation = false; // True if any batch has UV animation
+    uint8_t availableLODs = 0;  // Bitmask: bit N set if any batch has submeshLevel==N
 
     // Particle emitter data (kept from M2Model)
     std::vector<pipeline::M2ParticleEmitter> particleEmitters;
@@ -193,6 +194,7 @@ struct M2Instance {
 
     // Frame-skip optimization (update distant animations less frequently)
     uint8_t frameSkipCounter = 0;
+    bool bonesDirty = false;  // Set when bones recomputed, cleared after upload
 
     // Per-instance bone SSBO (double-buffered)
     ::VkBuffer boneBuffer[2] = {};
@@ -265,6 +267,8 @@ public:
     /**
      * Render all visible instances (Vulkan)
      */
+    /** Pre-allocate GPU resources (bone SSBOs, descriptors) on main thread before parallel render. */
+    void prepareRender(uint32_t frameIndex, const Camera& camera);
     void render(VkCommandBuffer cmd, VkDescriptorSet perFrameSet, const Camera& camera);
 
     /**