Tensor.prod returns a NaN when there is a negative value in the tensor

[quinton11]

Describe the bug
The prod and prod_dim tensor operation results contain NaN values if there is a negative value in the tensor performing the operation. This was tested with both ndarray and tch backends.

To Reproduce

use burn::{
    backend::{libtorch::LibTorchDevice, ndarray::{NdArray, NdArrayDevice}, LibTorch},
    tensor::{backend::Backend, Bool, Shape, Tensor, TensorData}};
    
    
    fn main() {
        type MyBackend = NdArray::<f32>;
        type MyTchBackend = LibTorch::<f32>;
        let device = NdArrayDevice::Cpu;
        let tch_device = LibTorchDevice::Cpu;

        // prod
        let nd_data = Tensor::<MyBackend,2>::from_data([[-1.0, 2.0, 3.0], [5.0, 9.0, 6.0]], &device);
        let nd_tensor = nd_data.clone().prod();
        println("{:?}",nd_tensor)
        // Tensor { primitive: Float(NdArrayTensor { array: [NaN], shape=[1], strides=[1], dynamic ndim=1 }) }

        // prod tch
        let tch_data = Tensor::<MyTchBackend,2>::from_data([[-1.0, 2.0, 3.0], [5.0, 9.0, 6.0]], &tch_device);
        let tch_tensor = tch_data.clone().prod();
        println("{:?}",tch_tensor)
        // Tensor { primitive: Float(TchTensor { tensor: [-1620], phantom: PhantomData<f32> }) }
    }

Expected behavior
Result should be [-1620] instead of NaN.

Screenshots

Desktop (please complete the following information):

• OS: [e.g. macos]
• Browser [e.g. chrome, safari]
• Version [Sequoia 15.0]

Additional context
The prod method calls the float_log implementation for the NdArray Backend which basically takes the natural log of the tensor's data, the result is then summed and the exponent taken.
Naturally the natural log of a negative number is undefined, explaining the NaN.
With pytorch however, negative numbers are catered for.

If this is desired behaviour this issue should be closed.

Edit: Not able to reproduce for the Tch Backend, NdArray however still returning NaN values