Pathfinder: add support on cutensorMp library loading (#2374)

* NEW: add support on cutensorMp library loading

* Fix cuTENSORMp CUDA runtime preload

CUDA 12's libcutensorMp.so.2 directly depends on libcudart.so.12, but its wheel RUNPATH does not include the cuda_runtime package. Pathfinder can therefore find and preload cuTENSOR, cuBLASLt, and NCCL yet still fail to dlopen cuTENSORMp in a wheel-only environment unless libcudart is already available from an ambient toolkit.

Add cudart to the descriptor dependency list so Pathfinder resolves the packaged CUDA runtime before opening cuTENSORMp. This makes CUDA 12 wheel-only loading independent of LD_LIBRARY_PATH while preserving CUDA 13 and local-toolkit behavior.

---------

Co-authored-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>

Author: yangcal

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/descriptor_catalog.py

@@ -375,6 +375,13 @@ class DescriptorSpec:
         site_packages_windows=("cutensor/bin",),
         dependencies=("cutensor", "cublasLt"),
     ),
+    DescriptorSpec(
+        name="cutensorMp",
+        packaged_with="other",
+        linux_sonames=("libcutensorMp.so.2",),
+        site_packages_linux=("cutensor/lib",),
+        dependencies=("cutensor", "cublasLt", "cudart", "nccl"),
+    ),
     DescriptorSpec(
         name="nccl",
         packaged_with="other",