Burn-WGPU tests fail on Windows with Radeon 6950

[hexd0t]

Describe the bug
Running run-checks.ps1 all fails for burn-wgpu with 4 failed tests during the std tests.

To Reproduce
Checkout the current main branch.
Open a Powershell Terminal, cd to the repo root, execute . .\run-checks.ps1 all

Expected behavior
All tests should pass.

Desktop (please complete the following information):

• OS: Windows
• Version: Server 2025 Preview (build 26212.5000), equivalent to Win 11
• GPU: Radeon 6950, Driver 31.0.24027.1012 / AMD Software 24.3.1
• rustc 1.78.0

Additional context

failures:

---- tests::jit::module_nearest_interpolate::tests::test_downsample_interpolation stdout ----
thread 'tests::jit::module_nearest_interpolate::tests::test_downsample_interpolation' panicked at crates\burn-wgpu\src\lib.rs:75:5:
Tensors are not approx eq:
  => Position 3: 7 != 6 | difference 1 > tolerance 0.0010000000000000002
  => Position 9: 161 != 160 | difference 1 > tolerance 0.0010000000000000002
  => Position 15: 315 != 314 | difference 1 > tolerance 0.0010000000000000002
  => Position 21: 469 != 468 | difference 1 > tolerance 0.0010000000000000002

---- tests::jit::remainder::tests::should_be_zero stdout ----
thread 'tests::jit::remainder::tests::should_be_zero' panicked at crates\burn-wgpu\src\lib.rs:75:5:
Tensors are not approx eq:
  => Position 0: 0 != 3.5 | difference 3.5 > tolerance 0.0010000000000000002
  => Position 1: 0 != 3.5 | difference 3.5 > tolerance 0.0010000000000000002
  => Position 2: 0 != 3.5 | difference 3.5 > tolerance 0.0010000000000000002

---- tests::jit_fusion::module_nearest_interpolate::tests::test_downsample_interpolation stdout ----
thread 'tests::jit_fusion::module_nearest_interpolate::tests::test_downsample_interpolation' panicked at crates\burn-wgpu\src\lib.rs:75:5:
Tensors are not approx eq:
  => Position 3: 7 != 6 | difference 1 > tolerance 0.0010000000000000002
  => Position 9: 161 != 160 | difference 1 > tolerance 0.0010000000000000002
  => Position 15: 315 != 314 | difference 1 > tolerance 0.0010000000000000002
  => Position 21: 469 != 468 | difference 1 > tolerance 0.0010000000000000002

---- tests::jit_fusion::remainder::tests::should_be_zero stdout ----
thread 'tests::jit_fusion::remainder::tests::should_be_zero' panicked at crates\burn-wgpu\src\lib.rs:75:5:
Tensors are not approx eq:
  => Position 0: 0 != 3.5 | difference 3.5 > tolerance 0.0010000000000000002
  => Position 1: 0 != 3.5 | difference 3.5 > tolerance 0.0010000000000000002
  => Position 2: 0 != 3.5 | difference 3.5 > tolerance 0.0010000000000000002


failures:
    tests::jit::module_nearest_interpolate::tests::test_downsample_interpolation
    tests::jit::remainder::tests::should_be_zero
    tests::jit_fusion::module_nearest_interpolate::tests::test_downsample_interpolation
    tests::jit_fusion::remainder::tests::should_be_zero

test result: FAILED. 1512 passed; 4 failed; 4 ignored; 0 measured; 0 filtered out; finished in 20.25s

[nathanielsimard]

I don't have an AMD GPU, but most of these are precision errors, where we should simply fix the test. The only exception is maybe the remainder operation in jit_fusion where I don't really understand the the 3.5 😅.