LpNorm + ScaleNorm

[ClashLuke]

ScaleNorm:

LayerNorm:

Otherwise, both models have identical settings, so purely by switching LayerNorm with ScaleNorm (both L2), the model becomes 25% faster while achieving the same (or better) convergence:

L1-Norm has the same speed but worse loss initially:

Additionally, this PR makes grad checks tougher to pass, as all models don't just get random parameters but also a random output gradient. This changed input ensures that we use the output gradient correctly, as our custom_grad functions could otherwise ignore it.

[ClashLuke]

All tests pass:

LayerNorm has a 3% higher loss than L2-ScaleNorm:

L1-ScaleNorm has an 8% higher loss than L2-ScaleNorm:

Regardless of that performance, I added an option to change the normalization's power and optionally centralize it.