Fix s1_train DDP crash on Windows single-GPU (sm_120 / Blackwell) - #2774

Summary

On Windows + single GPU + CUDA 12.8 + PyTorch 2.7+ on Blackwell (sm_120) hardware, s1_train.py crashes with a native access violation (exit code 3221225477 / 0xC0000005) shortly after pytorch_lightning's Trainer.fit() is called, before the first batch runs.

Repro

OS:      Windows 11
GPU:     RTX 5090 (compute capability 12.0, sm_120, Blackwell)
CUDA:    12.8
PyTorch: 2.11.0+cu128
Lightning: 2.6.1

Run any voice training through the standard webui / s1_train.py pipeline. The s1 (GPT) stage prints model summary, dataloader warnings, then exits with 3221225477 and no Python traceback.

Root cause

s1_train.py constructs the Trainer with DDPStrategy(backend="gloo") on Windows, regardless of GPU count. The gloo + sm_120 + CUDA 12.8 combination has a known incompatibility — confirmed by an independent PyTorch forum report titled "[Solved] RTX 5090 (sm_120) Training Segfault — DDP Was the Cause" — and produces a native crash inside Lightning's training-loop setup.

For a single GPU, DDP brings only process-management overhead and no parallelism benefit. Letting Lightning pick `strategy="auto"` selects `single_device`, avoiding DDP entirely.

Changes

Two narrowly-scoped patches, both Windows-conditional:

GPT_SoVITS/s1_train.py

• On Windows, use strategy=\"auto\" and devices=1 instead of DDPStrategy(\"gloo\") + devices=-1.
• Linux / macOS behaviour is unchanged — still NCCL DDP across all GPUs.

GPT_SoVITS/AR/data/bucket_sampler.py

• DistributedBucketSampler previously assumed a distributed group was always initialized. With strategy=\"auto\" on a single GPU, Lightning does not initialize the group, so dist.get_world_size() raises.
• Now: if the group is not initialized, fall back to a single-replica configuration (num_replicas=1, rank=0).
• Defensive change — behaviour is unchanged whenever DDP is initialized.

Testing

• Windows 11 + RTX 5090 (sm_120) + CUDA 12.8 + PyTorch 2.11+cu128 + Lightning 2.6.1
• 15-epoch v4 LoRA training on a 30-min Chinese dataset (`batch_size=4`, `lora_rank=32`)
• Pre-patch: crashes with `3221225477` after model load, no checkpoints saved.
• Post-patch: training completes cleanly through all 15 epochs, weights saved to `GPT_weights_v4/` as expected.

Closes

Closes #2626 ("`RuntimeError: makeDeviceForHostname(): unsupported gloo device`" — same root cause).