Add safe API for querying CUDA function attributes

[wizenink]

Adds support for querying CUDA kernel/function properties like register usage, memory requirements, and compilation info.

Changes

• Result layer: Added function::get_function_attribute() wrapper for cuFuncGetAttribute
• Safe API: Added CudaFunction::get_attribute() and convenience methods:
  • num_regs() - registers per thread
  • shared_size_bytes() - static shared memory size
  • const_size_bytes() - constant memory size
  • local_size_bytes() - local memory per thread
  • max_threads_per_block() - thread limit
  • ptx_version() / binary_version() - compilation versions
• Example: Added 10-function-attributes.rs demonstrating kernel introspection

Usage

let kernel = module.load_function("my_kernel")?;
println!("Registers: {}", kernel.num_regs()?);
println!("Shared mem: {} bytes", kernel.shared_size_bytes()?);

Useful for kernel optimization, debugging, and runtime capability checking.

[gemini-code-assist]

Code Review

This pull request introduces a safe API for querying CUDA function attributes, which is a valuable addition for kernel introspection. The implementation includes a low-level wrapper in result.rs, a high-level safe API in core.rs, and a clear example demonstrating the new functionality. The code is well-structured, but I've identified a significant thread-safety issue in the new CudaFunction::get_attribute method. It fails to bind the CUDA context to the calling thread, which can lead to errors in multi-threaded applications. I've provided a comment with a suggested fix to ensure correctness and thread-safety.

[wizenink]

Should now pass cargo fmt