Allow for covariant normalization in metatensor.learn.nn.EquiLayerNorm

[jwa7]

PR #513 is soon to be merged, which implements EquiLayerNorm as a metatensor-learn.nn building block. This maintains equivariance when applying a transformation to a TensorMap by only applying a LayerNorm() to the invariant blocks. For the covariants, an Identity() layer is applied. However, in principle normalisations can be applied to covariants iff they maintain covariance.

Opening this issue as feature request - extend the current EquiLayerNorm to allow for covariant normalization beyond just applying the Identity operation.