fix ratio

[hhaAndroid]

the loss calculation will convert logits to float32, so for alignment, we also need to convert it to float32 here to prevent the ratio from being 1 during rl training

[RangiLyu]

Suggested change

	# we also need to convert it to float32 here to prevent the ratio from being 1 during rl training
	# we also need to convert it to float32 to prevent the ratio from not being equal to 1 during on-policy rl training

[pppppM]

Due to the lack of use of chunk forward when computing old_logprobs, it may cause OOM (Out of Memory).
From the perspective of inference, the logits returned need to be bf16.
It is best to convert to float32 near the computation of logprobs.