feat(cuda): bundle Linux runtime libraries

[IdrisGit]

Summary

• Bundle libcudart, libcublas, and libcublasLt with Linux CUDA artifacts.
• Configure $ORIGIN RUNPATH so binaries load the bundled libraries.
• Disable unnecessary NCCL linkage.
• Verify packaged dependencies while keeping libcuda.so.1 driver-provided.
• Document that a current compatible NVIDIA driver is required; the CUDA Toolkit is not required.

Release coordination

This PR must be merged and released before the corresponding OpenWhispr application PR can ship: OpenWhispr/openwhispr#1982

After publishing the new whisper.cpp tag, the OpenWhispr PR must update WHISPER_CPP_TAG and the matching Linux CUDA artifact digest in src/helpers/whisperCudaManager.js.

Validation

• Built the Linux CUDA artifacts locally with CUDA 12.9.
• Verified RUNPATH and dependencies with readelf and ldd.
• Confirmed CUDA transcription through OpenWhispr using the packaged runtime.
• Confirmed CUDA → Vulkan fallback through OpenWhispr.

Before release

• Validate the unpacked CUDA artifacts with /usr/local/cuda-12.9 unavailable, ensuring no installed Toolkit libraries can hide a missing packaged dependency.
• Confirm libcudart, libcublas, and libcublasLt all resolve from the packaged artifact.
• Publish a new immutable tag rather than replacing an existing release asset.

Maintenance note

CUDA 12.9, CUDA major version 12, and /usr/local/cuda-12.9 currently appear in several workflow locations. A future cleanup can move these values into top-level GitHub Actions env: constants so CUDA upgrades require fewer coordinated edits. This is workflow maintenance only and does not require repository secrets or settings.

Supports OpenWhispr/openwhispr#1415.
Supports OpenWhispr/openwhispr#945.