fix: reject oversized cuTile partition dimensions

[Rachit2323]

Summary

• reject partition dimensions that do not fit the compiler and kernel i32 ABI
• propagate checked partition/map metadata conversion failures through normal launches and specialization-only warmup
• keep checked conversions at final kernel argument construction as defense in depth
• add driver-free boundary regressions for regular and mapped partitions

Fixes #1382.

Why

Partition and map dimensions enter the host API as usize, but the compiler metadata and kernel ABI use i32. The previous grid check allowed the full u32 range, while later as i32 casts changed values above i32::MAX into negative dimensions.

Testing

• cargo fmt -p cutile -p cutile-macro -- --check
• git diff --check
• cargo test -p cutile --lib partition_metadata_rejects_dimensions_above_i32_max could not run locally because this checkout has no CUDA 13+ toolkit; the build stopped in cuda-bindings/build.rs before compiling the changed crates

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.