fix: Pillow in PEP 723 deps + capture lighteval stderr on failure

Two fixes that landed after the first charon smoke run:

1. Pillow missing from PEP 723 deps
lighteval's task registry eagerly imports every task module at
startup (to build the discovery table). tasks/mathvista.py does
'from PIL import Image' at module load time. When uv builds a
clean venv from just eval.py's deps, PIL isn't pulled in — it
was only present on studio via some transitive. Charon's fresh
venv exposed the gap. Adding Pillow directly.

2. stderr silently swallowed in _run_model_rounds
The per-round lighteval subprocess was invoked with stdout+stderr
both → DEVNULL, so any crash (like the Pillow ImportError above)
produced no visible output and the caller just saw 'No round parquet
files produced' with no clue why. Now capturing stderr, checking
returncode, and printing the last 30 lines of stderr on failure.
stdout stays quiet (lighteval's noisy progress output is not useful
for success cases).

Found while bringing charon (first gguf worker) online — the charon
smoke surfaced both issues cleanly once we re-ran with the stderr fix.

Co-Authored-By: Virgil <virgil@lethean.io>

eval.py

@@ -10,6 +10,7 @@
 #     "pyarrow",
 #     "ruamel.yaml",
 #     "pyyaml",
+#     "Pillow",
 # ]
 # ///
 # SPDX-License-Identifier: EUPL-1.2
@@ -219,6 +220,11 @@ def _run_model_rounds(model_name, task, n_questions, rounds, tmp_dir, wrapper_fi
 
     wrapper_file is the absolute path to the LightevalModel subclass file
     (mlx_lm_wrapper.py or gguf_wrapper.py).
+
+    lighteval's stdout is quiet by default (noisy progress) but stderr is
+    captured and emitted if a round produces no output parquet — that makes
+    downstream "No round parquet files produced" failures debuggable
+    instead of silently swallowed.
     """
     import subprocess
     out_paths = []
@@ -240,7 +246,17 @@ def _run_model_rounds(model_name, task, n_questions, rounds, tmp_dir, wrapper_fi
                 "--output-dir", out_dir,
             ]
             print(f"  round {r}/{rounds} @ start={samples_start} for {model_name}", flush=True)
-            subprocess.run(cmd, check=False, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+            result = subprocess.run(
+                cmd,
+                check=False,
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.PIPE,
+                text=True,
+            )
+            if result.returncode != 0:
+                print(f"  lighteval exited {result.returncode} — stderr tail:", file=sys.stderr)
+                for line in (result.stderr or "").splitlines()[-30:]:
+                    print(f"    {line}", file=sys.stderr)
         parquets = glob.glob(f"{out_dir}/details/**/*.parquet", recursive=True)
         if parquets:
             out_paths.append(parquets[0])