NaN related ops are not correct for burn-wgpu fusion backend

[antimora]

I am adding two new tensor methods related to NaN (is_nan and contains_nan). My unit tests are failing for burn-wgpu backend. See #2088 PR

Describe the bug
These two methods do not work as expected:

    /// Returns a new tensor with boolean elements indicating whether each element of the input is NaN.
    ///
    /// # Returns
    ///
    /// A boolean tensor where `true` indicates NaN and `false` indicates a non-NaN value.
    pub fn is_nan(&self) -> Tensor<B, D, Bool> {
        // Check if the input tensor is NaN by comparing it to itself
        // NaN is the only value that is not equal to itself
        K::not_equal(self.primitive.clone(), self.primitive.clone())
    }

    /// Checks if the tensor contains any NaN values.
    ///
    /// # Returns
    ///
    /// A boolean tensor with a single element indicating whether the tensor contains any NaN values.
    pub fn contains_nan(&self) -> Tensor<B, 1, Bool> {
        // Summing the tensor will result in NaN if the tensor contains any NaN values
        // This is faster than checking each element individually
        // because it rolls up the NaN values into a single value
        let sum = K::sum(self.primitive.clone());

        // Check if the sum is NaN by comparing it to itself
        K::not_equal(sum.clone(), sum)
    }

To Reproduce

1. check out this PR branch: Add is_nan and contains_nan tensor ops #2088 PR
2. cd burn-wgpu
3. cargo test nan

Expected behavior
The tests should pass.

Screenshots

NOTE: The unit test should be enabled currently ignored for all backends.

[antimora]

@nathanielsimard @louisfd, this is only for fusion backend. Any particular difference?