Fix LPFormer attention dimensions

[BWAAEEEK]

This PR fixes LPFormer attention layer sizes for multi-layer and multi-head deployments.

The initial implementation makes reference to an undefined self.num_layers while constructing transformer layers. All of these changes fixed that, but multi-layer configurations still ended up producing mismatched pairwise feature dimensions because the final attention layer output consistently had 2 * hidden_channels size.

Also, multi-head attention output normalisation used out_channels instead of num_heads * out_channels.

Change Summary

This PR modifies the construction of LPFormer layers to use dimensions on edge feature input/output., and adds regression coverage for num_transformer_layers in {1, 2}, and num_heads in {1, 2}.

Tested with:

pytest test/nn/models/test_lpformer.py