Fix multi model ddp unused params

[Zish19]

I investigated issue #21548 and implemented a fix for unused parameter detection in MultiModelDDPStrategy.

Root cause:
MultiModelDDPStrategy wraps child modules in DistributedDataParallel, but returns the root LightningModule, so inherited DDPStrategy.pre_backward() exits early because self.model is not a DistributedDataParallel instance. As a result, prepare_for_backward() is never called and unused parameter detection is skipped.

Fix:
Override MultiModelDDPStrategy.pre_backward() and iterate over DDP-wrapped child modules. Only arm reducers for modules with active trainable parameters (requires_grad=True) to avoid false-positive reducer errors during multi-optimizer manual optimization (for example GAN training with toggle_optimizer()).

I also enabled a regression test for unused parameter detection in:
tests/tests_pytorch/strategies/test_multi_model_ddp.py

Branch:
Zish19:fix-multi-model-ddp-unused-params

Validation:

• test_multi_model_ddp.py
• test_ddp_integration.py

Both passed locally.