Optimization for ConvNd if dropout=0.

[gslama12]

As discussed in #2153

[BenjaminBossan]

Thanks for adding this DoRA optimization to conv layers. Generally, this looks good, but I have one comment about the bias term.

[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.

[BenjaminBossan]

Thanks for the update. Let's call make style or remove the trailing white space to make the linter happy.

[gslama12]

Should be fixed.

[BenjaminBossan]

Thanks for the update. As you can see, the tests are failing. I checked what's going on and the issue is that the bias is still flat at this point, hence when it is subtracted, PyTorch broadcasts the base_result.

By reshaping the bias, this should be addressed. IIUC, the bias shape should be (1, -1, ...), where ... are 1s, the number of which depends on the type of conv layer. I suggested a fix, but please LMK if you think this is incorrect.

[BenjaminBossan]

Thanks for bringing this DoRA optimization to conv layers.