Fix: with SAC, a new training batch should be sampled for each gradient step

[YiboDi]

…nt_step.

DDPG/TD3/SAC for robotics tasks #65
Increasing the number of parallel simulations accelerates the replay buffer data refresh rate in SAC and other off-policy algorithms. However, to fully leverage this increased data collection, the model should be updated more frequently. This can be achieved by increasing self._gradient_steps.

Issue:

- Current Behavior: The current implementation reuses the same training batch across multiple gradient steps, which can lead to overfitting and inefficient use of the new data collected from parallel simulations.

- Expected Behavior: According to the SAC algorithm implementation, a new training batch should be sampled for each gradient step to ensure diverse and fresh experiences are used for updates.

Fixed Implemented:

• Modified the SAC training loop to sample a new training batch from the replay buffer for each gradient step.