cubek-convolution: f16 1x1 conv returns wrong results when in_channels=96 and spatial positions >= 32768

[zachelnet]

Describe the bug

On the wgpu backend (Vulkan, AMD RADV/RDNA4) a f16 1x1 convolution returns
numerically wrong results (not just rounded) for a specific shape: input
channels = 96 and H*W >= 32768. The same conv is bit-exact against a f32
reference for every other shape tested.

To Reproduce

// burn 0.21 / burn-wgpu Vulkan, Backend = Vulkan<f16>
let mut conv = Conv2dConfig::new([96, 48], [1, 1]).init(&device);
// load any f16 weights + bias, e.g. random normal(0, 0.08)
let x = /* [1, 96, 128, 256] f16 input */;
let out = conv.forward(x); // wrong: mean abs err ~1.0 vs f32 reference

Shapes that reproduce (f16): [1, 96, 128, 256] (N=32768), [1, 96, 240, 320]
(N=76800), [1, 96, 256, 256] (N=65536).
Shapes that are correct: [1, 96, 128, 128] (N=16384), and [1, K, 240, 320]
for K in {48, 64, 80, 97, 112, 128}.

Expected behavior

f16 conv with f32 accumulation should match the f32 reference to ~1 ULP,
independent of the shape.

Environment

• burn / burn-wgpu / burn-cubecl 0.21.0, wgpu Vulkan on RADV
• GPU: AMD RX 9070 XT (gfx1201)