MRCuda: add getRuntimeInfo() returning Expected<RuntimeInfo>; deprecate isCudaAvailable()

[Fedr]

Summary

Introduces Cuda::getRuntimeInfo() as the primary way to query CUDA availability and device capabilities, and re-expresses the existing isCudaAvailable() on top of it (now deprecated).

The old isCudaAvailable() packed several concerns into one bool return plus four optional out-parameters: it queried the driver/runtime versions and the compute capability, and decided whether the device was suitable for computations — while discarding the underlying CUDA error on failure.

New API:

struct RuntimeInfo
{
    int driverVersion = 0;   ///< maximum driver supported version
    int runtimeVersion = 0;  ///< current runtime version
    int computeMajor = 0;    ///< compute capability major version
    int computeMinor = 0;    ///< compute capability minor version

    /// returns true if all versions pass the checks
    [[nodiscard]] MRCUDA_API bool fitForComputations() const;
};

/// Returns an error if CUDA is not available
MRCUDA_API Expected<RuntimeInfo> getRuntimeInfo();

getRuntimeInfo() returns the four versions on success, or a descriptive error string when no CUDA device is present or a CUDA call fails. The compute-capability-vs-runtime-version suitability logic now lives in RuntimeInfo::fitForComputations(), separate from the query.

Changes

• MRCudaBasic.h / .cpp — add RuntimeInfo + getRuntimeInfo(); propagate CUDA errors as Expected via the CUDA_RETURN_UNEXPECTED macro instead of returning a bare false.
• isCudaAvailable() — kept for backward compatibility (existing callers, the Python/C bindings and downstream code keep working unchanged), now marked [[deprecated( "Use getRuntimeInfo" )]] and reimplemented as a thin wrapper: it fills the optional out-parameters from getRuntimeInfo() and returns RuntimeInfo::fitForComputations().
• getCudaAvailableMemory() — no longer routes through the availability check; it tries cudaSetDevice( 0 ) directly and now also checks the result of cudaMemGetInfo (previously ignored), returning 0 on failure.
• Doc comments converted to /// Doxygen style.
• MRTestCuda — updated to the new API; logs the actual CUDA error message on failure and checks fitForComputations() explicitly.

Notes

• Backward compatible — isCudaAvailable() keeps its old signature and semantics (out-parameters + the fit-for-computations result), so no binding regeneration or downstream changes are required.