Feature 314: unify logging processing and add detailed global metrics for rollouts

[Sophie8]

1. This pr tries to implement [Feature Request] Improving logging processing logics #314 for the requests of detailed global metrics when sample mean for metrics like pg_loss, advantage in grpo can be nearly zero, we will gather global metrics as Max/Min/Std metrics to log more information for the users.
2. This pr also tried to address the issue of inconsistency of logging functions usage at different backend, tried to unify them under a same class called rolloutpostprocessor.py

sample rollout metrics:

{
"rollout/log_probs": -0.123456,
"rollout/ref_log_probs": -0.123456,
"rollout/rollout_log_probs": -0.125789,

"rollout/advantages_global_mean": 0.002300,
"rollout/advantages_global_std": 0.120000,
"rollout/advantages_global_min": -0.950000,
"rollout/advantages_global_max": 1.500000,

"rollout/values_global_mean": 0.010000,
"rollout/values_global_std": 0.450000,
"rollout/values_global_min": -3.200000,
"rollout/values_global_max": 2.800000,

"rollout/returns_global_mean": 1.230000,
"rollout/returns_global_std": 0.670000,
"rollout/returns_global_min": -0.200000,
"rollout/returns_global_max": 5.400000,

"rollout/entropy": 0.210000,
"rollout/response_lengths": 11.400000,

"rollout/step": 1210
}

[gemini-code-assist]

Important

Installation incomplete: to start using Gemini Code Assist, please ask the organization owner(s) to visit the Gemini Code Assist Admin Console and sign the Terms of Services.