Fixed GroupNorm implementation

[computer-whisperer]

• Confirmed that run-checks all script has been executed.
• Made sure the book is up to date with changes in this PR.

Related Issues/PRs

None

Changes

I discovered that GroupNorm was providing different outputs for the same input as the PyTorch version. The math error was relatively minor, but after this change it now matches the pytorch documentation, and the deviation I experienced in higher-level tests vs pytorch has been resolved.

I also had to tweak the expected test results in InstanceNorm slightly, as it uses GroupNorm internally.

Testing

The cargo tests for burn-core pass with this, and the observed higher-level deviation between my implementation of RWKV and the official pytorch implementation has been resolved.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.13%. Comparing base (8a6f37c) to head (404d60f).
Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2945      +/-   ##
==========================================
+ Coverage   82.11%   82.13%   +0.01%     
==========================================
  Files         867      867              
  Lines      120268   120268              
==========================================
+ Hits        98763    98778      +15     
+ Misses      21505    21490      -15     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[laggui]

Oh, good catch 😅

Epsilon should be added to avoid sqrt(0), looks like the previous implementation was incorrect.

Ref: https://paperswithcode.com/method/group-normalization