feat: ship with CUDA provider libs and preload CUDA runtime on load

[talmago]

Summary

• Bundle ORT CUDA provider plugins (providers_shared / providers_cuda) into CUDA-profile wheels only after auditwheel, so libonnxruntime_providers_cuda.so is found next to the vendored runtime (CPU wheels unchanged).
• Add nvidia-cublas (and onnxruntime-gpu[cuda,cudnn]) to the cuda extra so CUDA 13 user libs are installable via pip install "fast-lightonocr[cuda]".
• When runtime_kwargs selects execution_provider="cuda", call onnxruntime.preload_dlls(cuda=True, cudnn=True) before native session create so pip NVIDIA libs resolve without LD_LIBRARY_PATH. CPU loads skip this path entirely.
• Reorganize the Python README install / source-build docs (CPU vs CUDA).

Test plan

• Colab: BUILD_PROFILE=cuda pip install -v ".[cuda]" from this branch; wheel contains provider libs in fast_lightonocr.libs
• Colab: LightOnOCR.from_pretrained(..., runtime_kwargs={"execution_provider": "cuda"}) runs OCR
• Colab: nvidia-smi sampling during process() shows ~2 GiB / non-zero util
• Local: default CPU pip install -v ".[cpu]" still works (no provider bundling / no preload)
• Confirm BUILD_PROFILE=cpu wheel RECORD is unchanged vs pre-PR behavior