fix(build): add CUDA 13.2 support

[CharryWu]

Summary

• In CUDA 13.2, the id field of CUmemLocation_st was placed inside an anonymous union. Bindgen exposes this as __bindgen_anon_1: CUmemLocation_st__bindgen_ty_1 instead of a direct id field on the struct.
• The existing code (from commit 2623e21, "Fix CUDA 13 API issues") set id directly, which compiled on CUDA 13.0 but fails on 13.2.
• Adds a cuMemLocation_anon_id cfg in build.rs (gated on driver_version >= 13020) and uses field-level #[cfg] at all 5 CUmemLocation construction sites so both CUDA 13.0 and 13.2 compile correctly.

Test plan

• cargo build -p cust passes on CUDA 13.2 (Windows, RTX 3060, driver 595.79)
• Verify cargo build -p cust still passes on CUDA 13.0 (Windows, Linux CI)

Attribution

This bugfix was implemented with the help of Claude Sonnet 4.6 and Cursor IDE

[CharryWu]

@LegNeato Please hold off on this PR for now — Jimver/cuda-toolkit does not yet support CUDA 13.2.0 (Error: Version not available: 13.2.0), so the Windows CI matrix entry can't run, see CI output on my fork. I'll revisit once the action adds 13.2 support.

[LegNeato]

We have docker images for cuda 13 in https://github.com/Rust-GPU/rust-cuda/tree/e91b9cefc78258000f92d0978f327c907c515831/container for CI. They currently check 13.0...should we add 13.2? The matrix is starting to get big 🤔