Fix multi-GPU (DDP) training support

[u1vi]

Multi-GPU training with DDP strategy fails with AssertionError in dataset.py because prepare_data() only runs on rank 0 in DDP mode, leaving piper_config as None on other ranks.
This PR fixes four issues that prevent or affect multi-GPU training:

1. dataset.py — setup() crashes on non-rank-0 processes
   prepare_data() only runs on rank 0 in DDP mode (by Lightning design), but it's where self.piper_config gets set. setup() runs on all ranks and asserts self.piper_config is not None, which fails on ranks 1+. Fix: load the config from the JSON file that prepare_data() already writes to disk.
2. lightning.py — Missing toggle_optimizer/untoggle_optimizer
   With manual optimization + multiple optimizers in DDP, toggle_optimizer is needed so DDP correctly synchronizes only the relevant gradients during each backward pass. This also makes retain_graph=True unnecessary since the generator and discriminator loss graphs are already independent (via y_hat.detach()).
3. lightning.py — LR schedulers never stepped
   With automatic_optimization = False, Lightning does not automatically step LR schedulers. Added on_train_epoch_end to step them manually. Note: this was also a bug on single-GPU training — the learning rate decay was never being applied.
4. main.py — Auto-detect multi-GPU and set DDP strategy
   VITS has conditional code paths (e.g., stochastic duration predictor, speaker embeddings) that leave some parameters unused in certain steps. DDP requires find_unused_parameters=True to handle this. Instead of requiring users to pass an extra flag, multi-GPU is now auto-detected and the correct strategy is set automatically.
   Usage:
   No extra flags needed — multi-GPU training works out of the box

Tested with: 4x GPU setup, PyTorch Lightning, Python 3.13