feat(training): enable LyCORIS 4 fused kernels

[DualChimerra]

Summary

• upgrade LyCORIS to 4.0.0 and pair triton-windows with the installed PyTorch minor
• route plain T-LoRA through the official fused bypass API without materializing a full delta weight
• preserve the LoKr CUDA rank-dropout fix while delegating weight construction to upstream LyCORIS
• add kernel diagnostics and a cleaner Windows launcher banner
• document the integration decision and fallback behavior

Validation

• 132 focused LyCORIS, adapter, inference, and launcher tests passed
• RTX 4070 Ti CUDA smoke test: fused T-LoRA forward/backward completed with finite gradients
• direct Triton vs eager FP16 forward/backward parity passed
• Studio started successfully with LyCORIS 4.0.0 / Triton 3.6 selected; previous xformers/Triton warning is gone
• pip check: no broken requirements
• full suite: 3444 passed; 6 unrelated existing/environment failures (2 CUDA OOM block-swap cases on 12 GB, 4 platform/localization expectations)