Speed up synthetic generator hot paths by removing avoidable quadratic and allocation-heavy work

[Copilot]

Synthetic generators had avoidable overhead in core sampling loops, especially in social/work edge generation and multiplex mapping accumulation. This PR makes targeted internal optimizations to reduce generator runtime without changing public APIs or output semantics.

• Work-layer generation: eliminate cross-department pair checks

  • In make_social_network(), replaced global O(n^2) pair scanning with department-local pairing.
  • Effect: only evaluates candidate pairs that can actually produce work edges.

• Multiplex generator: remove repeated list reallocations

  • In random_multiplex_generator(), replaced dict.get(...)+[x] list concatenation with in-place append via setdefault(...).append(...).
  • Effect: avoids per-iteration list copying in two accumulation paths (layer_to_nodes, edge_to_layers).

• Minor cleanup

  • Removed an unnecessary cast in the department grouping path and a small typing/literal cleanup in random_generators.py.

Example of the core optimization pattern:

# before
layer_to_nodes[l] = layer_to_nodes.get(l, []) + [node]

# after
layer_to_nodes.setdefault(l, []).append(node)