Use 0.4 for depth unorm value in depth_compare_func rendering test (#4374)

This PR make depth_compare_func rendering operation test use 0.4 for
depth unrom value instead of 0.5001, which is exactly representable
in both depth16unorm and depth24unorm, and closely approximated
in depth32float.
This can help prevent shaders and depthClearValue get rounded in
different way, and make Qualcomm devices pass the test when using
DXGI_FORMAT_D24_UNORM_S8_UINT format for depth24plus-stencil8.

* Update src/webgpu/api/operation/rendering/depth.spec.ts

Co-authored-by: Kai Ninomiya <[email protected]>

---------

Co-authored-by: Kai Ninomiya <[email protected]>

Author: jiangzhaoming

src/webgpu/api/operation/rendering/depth.spec.ts

@@ -289,9 +289,12 @@ g.test('depth_test_fail')
     t.runDepthStateTest(testStates, _expectedColor);
   });
 
-// Use a depth value that's not exactly 0.5 because it is exactly between two depth16unorm value and
-// can get rounded either way (and a different way between shaders and clearDepthValue).
-const kMiddleDepthValue = 0.5001;
+// Use a depth value of 0.4, which is exactly representable in depth16unorm (26214 / (2^16-1))
+// and depth24unorm (6710886 / (2^24-1)), and closely approximated in depth32float
+// (0.4000000059604644775390625).
+// This can help prevent shaders and depthClearValue get rounded in different way making equal
+// comparison result unexpected.
+const kMiddleDepthValue = 0.4;
 
 g.test('depth_compare_func')
   .desc(