Add counter-target selectivity

[ssiddhantsharma]

Adds counter-target selectivity to the loop: alongside the on-target ligand, each
round's designs are also folded against one or more off-target ligands, and two
objectives reward on-target binding minus the worst-case off-target binding:

• iptm_selectivity — on_iptm - max(offtarget_iptm)
• pbind_selectivity — on_affinity_probability - max(offtarget_affinity_probability)

This is the paired on-minus-off difference: a design that binds the target but not
the counter-target scores highest, so you can steer designs away from a similar
molecule you don't want bound.

Off by default; enabled with --counter-target-smiles <SMILES> [...] (or
counter_target_smiles=[...]).

How it fits

Same seams as the existing objectives, no restructuring:

• fold_counter_targets() folds each design against each counter-target using the
  existing predict_complex_structures() / get_boltz_yaml_boilerplate(), writes to
  its own offtarget_<i>_boltz_inputs dir, and keeps the worst-case (strongest)
  off-target binding per design keyed by Boltz stem. Off-target structures are only
  scored, never promoted to backbones.
• parse_boltz_binding_metrics() reads iptm (+ affinity_probability_binary when
  predicting affinity) from the Boltz outputs.
• identify_backbone_candidates() injects offtarget_* into the existing
  confidence_data dict, so compute_objective_function() just sees new keys.

Wired through run_nise_boltz2x_cli.py too (--counter-target-smiles plus the new
objectives in the choices list).

Notes

• Multiple counter-targets take the worst case via nanmax, so a design that binds
  any of them is penalised, and a missing off-target affinity is skipped rather than
  propagated as nan.
• The on-minus-off difference doesn't put a floor on on-target binding on its own —
  combine it with an on-target objective/threshold if you want that.
• Cost: one extra fold per counter-target per round.

Testing

• tests/test_selectivity.py covers the objectives (including a case where a
  selective design beats a promiscuous one) and the Boltz metric parser.
• Checked against real Boltz output that the off-target fold lands where the code
  expects and parses correctly; the fold reuses the same predict path NISE already
  runs for the on-target ligand.