feat: add Low GPU Accuracy setting for maximum performance

Implements a new "Low" GPU accuracy level that prioritizes performance
over accuracy by aggressively cutting corners in GPU emulation.

Changes:
- Add GpuAccuracy::Low enum and setting infrastructure
- Implement IsGPULevelNormal() helper function
- Skip texture cache checks and query operations
- Use unsafe memory reads for DMA operations
- Disable fence delays and query precision
- Add UI support for desktop (Qt) and Android

Performance optimizations:
- Skips texture cache coherency checks (vk/gl_rasterizer.cpp)
- Non-blocking query synchronization (query_cache.h)
- Unsafe memory operations (dma_pusher.cpp)
- No macro parameter refresh (maxwell_3d.cpp)
- Immediate fence signaling (fence_manager.h)
- Non-precise Vulkan queries (vk_query_cache.cpp)

Ideal for lower-end hardware and users prioritizing FPS over accuracy.
Works on both desktop and Android platforms.

Signed-off-by: Zephyron <zephyron@citron-emu.org>

src/video_core/dma_pusher.cpp

@@ -98,7 +98,8 @@ bool DmaPusher::Step() {
                         &command_headers);
             ProcessCommands(headers);
         };
-        if (Settings::IsGPULevelHigh()) {
+        if (Settings::IsGPULevelNormal()) {
+            // Normal/High/Extreme: Use safe reads for most operations
             if (dma_state.method >= MacroRegistersStart) {
                 unsafe_process();
                 return true;
@@ -106,6 +107,10 @@ bool DmaPusher::Step() {
             safe_process();
             return true;
         }
+        // Low accuracy: Use unsafe reads for maximum performance everywhere
+        unsafe_process();
+        return true;
+        // Note: The code below is unreachable for Low, but kept for reference
         // Even in normal accuracy, use safe reads for KeplerCompute inline methods
         if (subchannel_type[dma_state.subchannel] == Engines::EngineTypes::KeplerCompute &&
             dma_state.method == ComputeInline) {