fix: smaller per-device batch + grad accum + bigger completion budget for L40S

fix: pass --per-device-train-batch-size 4 to GRPO so effective batch divides num_generations

add: pre/post eval + summarize + bumped GRPO config (space/training/app.py)

fix: multi-GPU SFT shape mismatch (space/training/app.py)

initial: drugenv trainer control panel