Add CUDA distributed training runtime

[antlobach]

Summary

• upgrade JavaCPP PyTorch/CUDA presets and expose CUDA device controls
• add NCCL collectives, managed rank launch, DDP, distributed sampling, AMP, and rank-zero checkpoints
• commit reduced gradient buckets only after NCCL completion so every model parameter stays synchronized
• document Java, Clojure, CUDA 13.1, cuDNN, NCCL, multi-GPU setup, launch, training, and validation
• test the CPU suite across Java 21, 22, 23, 24, and 25

Verification

• Linux Java 25 CPU release suite: 163 tests, 722 assertions, 0 failures, 0 errors
• dual RTX A5000 GPU release suite: 163 tests, 722 assertions, 0 failures, 0 errors
• CUDA smoke: two-rank NCCL/DDP, all-parameter synchronization, AMP, fused attention, checkpoint restore, worker failure propagation, and process cleanup passed
• deterministic two-rank Clorch/PyTorch comparison: 12 losses and trained parameters were float32-exact; maximum decimal error was 1.4984131269102363E-8
• clj-kondo MCP: 0 errors and 0 warnings in changed Clojure files
• MkDocs strict build passed