Feat: skip nvidia CUDA wheels when system provides compatible libs

Detects system CUDA installation and version. Skips downloading nvidia-*
wheel packages (~2.8GB) when the system CUDA version >= what the wheels
require. Each wheel is checked individually — only skipped if the system
has both the correct library AND a compatible CUDA version.

For RunPod's CUDA 12.4, this correctly determines the wheels need CUDA 12.8
and downloads them anyway (no ABI mismatch). On newer systems with CUDA 12.8+,
this would save ~2.8GB of downloads and cut cold start nearly in half.

Writes stub .dist-info for skipped packages so importlib.metadata finds them.
Sets LD_LIBRARY_PATH for system CUDA when launching Python.

Opt-out: ZS_NO_SYSTEM_CUDA=1

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: lightsofapollo