Fix CUDA init races and add safety checks

[sethuiyer]

Motivation

• Prevent nondeterministic behavior and hard-to-debug races in the CUDA Ramsey kernel initialization by removing concurrent RNG usage across threads.
• Fail fast on invalid host-side launch parameters to avoid undefined GPU behavior for unsupported state sizes or zero/negative grids.
• Make the sample GPU host code more robust by surfacing allocation and copy failures immediately.

Description

• Replaced per-thread cooperative RNG initialization in ramsey_baha_kernel_optimized with a single-thread (tid == 0) initialization loop to avoid races and then __syncthreads(); added #include <climits> for INT_MAX usage (file: include/baha/baha.cuh).
• Added input validation checks in launch_ramsey_optimizer to return cudaErrorInvalidValue when num_blocks <= 0 or words_per_state is out of the supported range (file: include/baha/baha.cuh).
• Hardened the GPU demo (src/baha_gpu.cu) by adding missing standard headers (<cmath>, <ctime>) and replacing raw cudaMalloc/cudaMemcpy/cudaFree calls with the existing cudaCheckError wrapper so allocation/copy/free errors are detected and reported.

Testing

• No automated tests were run for this change.

---

Codex Task