config: default filter_uniform_groups=true, refill_filtered_uniform_groups=false

Flip the rejection_sample defaults in the base config so uniform (zero-advantage)
groups are filtered out and NOT refilled — i.e. dropped from the step rather
than kept as zero-advantage placeholders. base.yaml is the sole definer of these
keys; every backend/cookbook config inherits it.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: jeffreysijuntan

rllm/trainer/config/rllm/base.yaml

@@ -137,8 +137,8 @@ rejection_sample:
   multiplier: 1
   min_partial_solve_tasks: 1
   min_trajs_per_group: 2
-  filter_uniform_groups: false
-  refill_filtered_uniform_groups: true
+  filter_uniform_groups: true
+  refill_filtered_uniform_groups: false
 
 # Remote Runtime Configuration (for agent-in-container runtimes like AgentCore, Harbor)
 remote_runtime: