fix: correctly handle the inconsistency in reward_extra_info and reward_tensor

[0x404]

If we are using multiple datasources for training/test, some reward function may return float and some may return dict (including extra_info) depending on the datasource.

In this case, the reward_tensor and reward_extra_info returned by RewardManager will have different shape, this causes the error in #1031.

this PR fix the inconsistency by correctly padding a place holder 'unkown' for those reward function only return float, this place holder will not be processed in and has no effect on the correctness:

verl/verl/trainer/ppo/metric_utils.py

Lines 227 to 233 in dc1714a

	# Calculate metrics for each group
	data_src2prompt2var2metric = defaultdict(lambda: defaultdict(lambda: defaultdict(dict)))
	for data_source, prompt2var2vals in data_src2prompt2var2vals.items():
	for prompt, var2vals in prompt2var2vals.items():
	for var_name, var_vals in var2vals.items():
	if isinstance(var_vals[0], str):
	continue

[0x404]

Hi, @vermouth1992, Could you please help review this PR? Much appreciated!