fix(rpm): filter auto-detected CUDA library dependencies

The binary's ELF headers cause rpm to auto-require exact soname
versions of CUDA libraries from the build host, forcing downgrades
on systems with newer compatible versions. Filter these out with
__requires_exclude so consumers can use any compatible CUDA install.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

rpm/mistralrs.spec

@@ -21,11 +21,12 @@ Source2:        mistralrs@.conf.example
 
 ExclusiveArch:  x86_64
 
-# Runtime requirements. We link against the CUDA runtime; consumers must have
-# a matching CUDA installation or the rpmfusion nvidia driver's cuda-libs.
-# We don't hard-require it at the RPM level because consumers may have CUDA
-# from multiple sources (nvidia direct, rpmfusion, etc.) — failing to load
-# libcuda.so at runtime gives a clearer error than RPM dep resolution would.
+# Suppress auto-detected CUDA library dependencies. The binary links against
+# the CUDA runtime, cuDNN, NCCL, etc. but we don't want rpm to pin exact
+# soname versions — consumers may have CUDA from multiple sources (nvidia
+# direct, rpmfusion, etc.) and different compatible versions. A runtime
+# dlopen failure gives a clearer error than rpm dep resolution would.
+%global __requires_exclude ^lib(cuda|cudart|cudnn|cublas|cublasLt|curand|nvrtc|nccl)
 Requires:       systemd
 
 # Flavours are mutually exclusive with other flavours of themselves at the