add v5-S2 3-seed ensemble ckpts (seeds 1024/5555/789) — canonical Q_theta from paper (rho_bar = 0.5201 on 8 OOD)

Per-seed OOD test rho_bar (TS-S2 8-target split cam/bcl2/era/mdm2/ran/a2a/pai1/integrin):
seed 1024: 0.5298
seed 5555: 0.5200
seed 789 : 0.5106
ensemble mean: 0.5201

These replace the older single-seed `checkpoints/Q_theta_phase2.pt` (v2 strict holdout, ~3.6 MB, 897k params) for reproducing paper numbers with rescore.py. The v2 checkpoint is left in place for backward compatibility but is not canonical — v5-S2 is bigger arch (~2.8M params, ~10.4 MB) and pooled-training on 51 train targets.

checkpoints/v5s2/seed_1024/best_phase2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fb885db114c3b231a962be0bdb0900d69515f903cf0bb526777f0f75c54be92
+size 10862978

checkpoints/v5s2/seed_5555/best_phase2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4e21949a988336d2fc20171186d94da8143512ba707defc837f242ce8a13611
+size 10862978

checkpoints/v5s2/seed_789/best_phase2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dbcf9a3f2c9092ad01d6c71f70efa871d748a02f2e8b14ba6e31ebba78455cd
+size 10862978