feat(native): Add cuBLAS dynamic loader (alongside cuBLASLt)

[m96-chan]

Overview

Create a dynamic loader for cuBLAS, similar to the existing cuBLASLt loader.

Current State

native/jit/
├── cublaslt_loader.cpp   (1080 lines, cuBLASLt dynamic loading)
├── cublaslt_loader.hpp
├── nvrtc_loader.cpp
└── nvrtc_loader.hpp

cuBLASLt is loaded dynamically, but cuBLAS is not available.

Proposed Addition

native/jit/
├── cublas_loader.cpp     (NEW - cuBLAS dynamic loading)
├── cublas_loader.hpp     (NEW)
├── cublaslt_loader.cpp
├── cublaslt_loader.hpp
├── nvrtc_loader.cpp
└── nvrtc_loader.hpp

Required Functions

Core BLAS Operations

// GEMM
cublasSgemm()   // FP32
cublasDgemm()   // FP64
cublasHgemm()   // FP16
cublasGemmEx()  // Mixed precision

// GEMV
cublasSgemv()
cublasDgemv()

// Handle management
cublasCreate()
cublasDestroy()
cublasSetStream()
cublasSetMathMode()

Dynamic Loading Pattern

// Follow same pattern as cublaslt_loader
class CuBLASLoader {
public:
    static CuBLASLoader& instance();
    bool is_available() const;
    
    // Function pointers
    cublasStatus_t (*sgemm)(...);
    cublasStatus_t (*dgemm)(...);
    cublasStatus_t (*hgemm)(...);
    cublasStatus_t (*gemmEx)(...);
    // ...
    
private:
    void* handle_ = nullptr;
    bool load_library();
};

DLL Search Order (Windows)

cublas64_13.dll  // CUDA 13.x
cublas64_12.dll  // CUDA 12.x
cublas64_11.dll  // CUDA 11.x

Use Cases

1. Fallback path - When custom kernels fail, use cuBLAS
2. Correctness reference - Compare custom kernel results vs cuBLAS
3. Legacy dtype support - FP64 GEMM via cuBLAS
4. Batched GEMM - cublasSgemmBatched for small batch sizes

Acceptance Criteria

• Dynamic loading works on Windows/Linux
• Graceful fallback if cuBLAS not found
• Handle caching (single instance)
• Basic GEMM operations exposed
• Python binding for availability check

Related

• Existing: native/jit/cublaslt_loader.cpp
• refactor(native): Split ops_bindings.cpp into modular files #131 (binding split) - Add to appropriate binding file