bdh.py: add extend_freqs() for phase-preserving latent-width growth

[asb-42]

Problem

RoPE frequencies are exponent-normalized by the neuron count (theta ** (-i/n) in
get_freqs). Any code that grows the latent width (progressive networks, upsizing
experiments, model editing) and simply recomputes get_freqs(n_new) therefore
**silently rewrites the phases of every existing neuron** and corrupts the trained model —
even though all actual weights are untouched.

We hit this while building a continual-learning growth mechanism on top of \BDH{}: a grown
model evaluated zero steps after width extension already behaved like an untrained model,
and the root cause was exactly this frequency-table renormalization.

Fix

extend_freqs(freqs, n_new) keeps the existing frequency table verbatim and appends
entries for new neurons computed at the new scale. Width growth becomes phase-preserving for
all pre-existing neurons by construction.

Verification

New test_bdh.py covers:

• extended tables preserve the prefix exactly (assert_close);
• naive recomputation at the larger count demonstrably changes the prefix (documents the pitfall);
• Attention builds with an extended table.

All three pass.

Scope

Deliberately minimal: one helper + docstring warning on get_freqs + tests. A full
continual-learning suite (phase-ordered training, growth/routing/consolidation mechanisms)
is being maintained separately in asb-42/bdh and can be
discussed separately if of interest.