Drop batch size to 16 (32 OOMs on 5090 with PGNet's variable input size)

PGNet uses dynamic image sizes per batch (max_text_size=512), which spikes
peak memory above 32GB at batch 32. Settle on 16 — ~2x throughput vs the
default 14 while staying well clear of OOM.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

configs/kr_lp_pgnet.yml

@@ -83,7 +83,7 @@ Train:
           channel_first: False
       - E2ELabelEncodeTrain:
       - PGProcessTrain:
-          batch_size: 32
+          batch_size: 16
           use_resize: True
           use_random_crop: False
           min_crop_size: 24
@@ -95,7 +95,7 @@ Train:
   loader:
     shuffle: True
     drop_last: True
-    batch_size_per_card: 32
+    batch_size_per_card: 16
     num_workers: 8
 
 Eval: