Implement numerically stable log_softmax()

[taha-yassine]

A numerically stable implementation of log_softmax(). It's similar to the current implementation of softmax() and is what's used in PyTorch and SciPy too.

log_softmax([1000,1])

// Current implementation
// > Array [ 0, -Infinity ]

// Proposed implementation
// > Array [ 0, -999 ]

[xenova]

Thanks so much! 🤗 Can you add a unit test for this to tests/maths.test.js? One normal case and one case where the unstable version will break? 🔥

[taha-yassine]

Done :)

[xenova]

Thanks so much! Approved!

As seen in #667, a standard for loop is faster for computing sums, so I've updated that. I tested both and indeed it is faster:

(code block 1 = reduce; code block 2 = normal for loop)

It also removes the need for @ts-ignore

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[xenova]

(ignore failing tests; related to something else)