[WIP] Improve variance scaling and reduce perplexity in transformers

[Copilot]

• Analyzed transformer v1 and v2 implementations
• Fix MhaResidualNode.forward() in transformer_v2.py:
  • Add causal context scaling for softmax averaging (sqrt(eff_ctx))
  • Add 1/√2 residual scaling to prevent variance doubling
• Fix Mlp2ResidualNode.forward() in transformer_v2.py:
  • Add 1/√2 residual scaling to prevent variance doubling
• Fix create_deep_transformer() in transformer_v2.py:
  • Use proper per-node initialization (dimension-appropriate init for each node type instead of a single global std)
• Fix SkipConnection.forward() in skip_connection.py:
  • Add 1/√N scaling for N inputs (prevents variance doubling in v1 transformer with external skip connections)
• Add variance tests for v2 transformer nodes
• Run existing tests to confirm no regressions