fix(gpu): support CUDA 13.x toolkits (CUmemLocation anon-union)

[cynd22]

What

Building ferrotorch-gpu against a CUDA 13.x toolkit (CUDARC_CUDA_VERSION=13020) fails to compile:

error[E0609]: no field `id` on type `CUmemLocation_st`
   --> ferrotorch-gpu/src/graph.rs:149
        props.location.id = device;

In CUDA 13.x, cudarc's CUmemLocation moved id into an anonymous union (__bindgen_anon_1.id). The value semantics are identical. This is the only thing blocking a CUDA-13 build — the rest of the crate compiles cleanly at 13020.

Fix

• graph.rs: the single assignment is cfg-gated — props.location.id on CUDA 12.x, props.location.__bindgen_anon_1.id on CUDA 13.x.
• build.rs: emits a new ferrotorch_cuda13 cfg when the resolved CUDA version is >= 13000 (reads CUDARC_CUDA_VERSION, falling back to an nvcc --version probe), plus the matching rustc-check-cfg.

The default 12080 build (pinned in .cargo/config.toml) never sets the cfg, so the 12.x path is byte-for-byte unchanged.

Verification

On a CUDA 13.2 / RTX 2070 Super host (Arch, driver 595.71, nvidia-open):

• cargo build -p ferrotorch-gpu --features cuda — clean at both CUDARC_CUDA_VERSION=12080 and 13020.
• At 13020, runs natively with no CUDA-12 runtime shim (system /opt/cuda cuBLAS/cudart .so.13): init_cuda_backend() succeeds, and a small MLP plus a GPT-2-small-shaped transformer (cuBLAS GEMM, LayerNorm, attention/softmax, GELU) produce outputs matching a PyTorch fp32 reference to max_abs ~= 8e-6.

Note / scope

This covers the cuBLAS + elementwise/attention path. cuSOLVER / cuFFT / cuSPARSE under a 13.x pin still expect their .so.13 libraries to be present — orthogonal to this change, just calling it out so the scope is clear.