Tensor::log_sum_exp() loses precision for small numbers

[leejuyuu]

The current implementation of the Tensor::log_sum_exp() function does not handle small numbers correctly. For example, the following code

let arr = Tensor::full(-1000_f64, (10,), &Device::Cpu)?;
let logsumexp = arr.log_sum_exp(0)?;
println!("{:?}", logsumexp);

prints

Tensor[-inf; f64]

whereas on torch 2.3.0

>>> arr = torch.tensor([-1000]*10)
>>> torch.logsumexp(arr, 0)
tensor(-997.6974)

Version info

candle-core = "0.6.0"
Tested on Arch Linux

[LaurentMazare]

Closing based on the fix in #2367, feel free to reopen if it's still an issue.