[BUG] Memory leak in loss backward when enabled torch activation checkpoint & zero3 & single-model-multi-branch network

Describe the bug
Huge memory leak in loss backward when enabled torch activation checkpoint & zero3 & single-model-multi-branch network.

To Reproduce
Steps to reproduce the behavior:
Training any single-model-multi-branch network with torch activation checkpoint & zero3 optimizer, e.g training diffsynth's animate-v2-lora model

1. git clone https://github.com/modelscope/diffsynth-studio
2. update examples/wanvideo/model_training/lora/Wan2.2-Animate-2-14B.sh add --config_file examples/wanvideo/model_training/full/accelerate_config_zero3.yaml to the launch commad
3. running: sh examples/wanvideo/model_training/lora/Wan2.2-Animate-2-14B.sh
4. and the GPU Memory will increase quickly during the loss backward
   Expected behavior
   Should release the zero3 all-gathered parameters during loss backward even AC is enabled and recompute happened

Screenshots

System info (please complete the following information):

• OS: [Ubuntu 22.04]
• GPU count and types [single machines with x8 32G RTX5090]
• Python version: 3.10.14

Launcher context
accelerate launch with deepspeed zero3 optimizer

Docker context
No Docker context

Additional context
After some debug and text, I got the memory leak is caused by the wrong ds_grads_remaining reset and count
in deepspeed/runtime/zero/parameter_offload.py's PostBackwardFunctionModule hook，
In AC & zero3 & single-model-multi-branch network env， the PostBackwardFunctionModule would be registered double times but the ds_grads_remaining is reset by _run_after_backward_function in every hook register so post_backward_function only run once, but single-model-multi-branch network should run multiple times depends on the number of branches, so the params gathered by pre-hook would never be released.

Start with deepspeed/runtime/zero/parameter_offload.py, focusing on the PostBackwardFunctionModule hook, _run_after_backward_function, and ds_grads_remaining. Reproduce the issue with the diffsynth-studio Wan2.2-Animate-2-14B.sh command and the stated zero3 configuration. Done means gathered parameters are released during loss backward across all branches when activation checkpointing is enabled, without continued GPU memory growth.