Show every distinct GRPO task type in the first N rows (#238)

Previously the GRPO section used a round-robin interleave that split
rows into a vLLM bucket and a non-vLLM bucket and walked each bucket
in turn. That guaranteed uniqueness within each bucket, but not across
the whole section: the top 10 rows were all vLLM (with GSM8K and DAPO
repeating because the vLLM bucket only had four distinct types),
while the first non-vLLM distinct type (Sudoku) did not show up until
row 11. Duplicates were also grouped by bucket instead of ordered by
popularity across the whole section.

Replace _interleave_by_task with _unique_types_first:

1. Sort every GRPO row by (popularity, count_ok, has_vllm) descending
   so the tuple ordering is popularity first, then vLLM as a minor
   tiebreaker between rows of the same popularity and task type.
2. Walk the sorted list once. The first time a task_type is seen,
   that row is the representative of that type; every subsequent row
   with the same task_type is a duplicate.
3. Emit all representatives first (already in popularity order from
   step 1), then all duplicates (also in popularity order).

Result for the current GRPO pool (11 distinct task types):

  Row 1   GSM8K Math + vLLM    (Llama3.1 8B)
  Row 2   Sudoku               (NeMo Gym Sudoku)
  Row 3   Multi Environment    (NeMo Gym Multi Environment)
  Row 4   2048 Game            (gpt oss BF16 20B)
  Row 5   Minesweeper Game     (gpt oss 20B)
  Row 6   Auto Kernel Creation (gpt oss 20B)
  Row 7   DAPO Math + vLLM     (Qwen3 8B FP8)
  Row 8   ORPO                 (Llama3 8B)
  Row 9   Wordle + vLLM        (Openenv wordle)
  Row 10  Vision Math + vLLM   (Qwen2.5 VL 7B)
  Row 11  DPO                  (Zephyr 7B)
  Row 12..29  duplicates in popularity order

Rows 1-11 now cover every distinct task type in the section, which
was the stated goal. Non-GRPO sections are untouched.

Author: danielhanchen