fix(grpo): fix list/list division TypeError in partial parse reward calculation

[isaacbmiller]

When computing the format reward for partially-parsed outputs during GRPO
bootstrapping, the code divides two lists (present / expected) instead of
their lengths. This raises TypeError: unsupported operand type(s) for /: 'list' and 'list'.

The intent is to compute the fraction of expected output fields that were
successfully parsed, so len(present) / len(expected) is the correct
expression.