[ILUVATAR] Support eig

[HydrogenSulfate]

This pull request adds support for the MAGMA library to the Iluvatar GPU backend in Paddle, enabling efficient computation of eigenvalues and eigenvectors (eig) for matrices. It introduces new CMake configuration for MAGMA, updates build scripts to include and link MAGMA, adds a CUDA eig kernel implementation leveraging MAGMA, and provides comprehensive unit tests for the new eig operator.

Build system integration and dependencies:

• Added external/magma.cmake to set up MAGMA as a precompiled third-party library, including download, verification, and build steps.
• Updated backends/iluvatar_gpu/CMakeLists.txt to include MAGMA in the build process, add relevant source files (magma_function.cc, magma.cc), and link the MAGMA library. [1] [2] [3]

Eigenvalue kernel and runtime support:

• Implemented the CUDA eig kernel in eig_kernel.cu, using MAGMA for computation and supporting both real and complex input types. Handles CPU-GPU data movement as required by MAGMA.
• Extended the runtime context in iluvatar_context.h to manage a cuSolver handle, required for MAGMA operations. [1] [2]

Testing and validation:

• Added a comprehensive unit test suite for the eig operator on Iluvatar GPU, covering correctness, gradients, error handling, and both static and dynamic graph modes.

Submodule update:

• Updated the Paddle submodule to reference a new commit, likely incorporating upstream changes needed for MAGMA integration.

[paddle-bot]

Thanks for your contribution!