Refactor CUDA evaluation functions and improve multi-GPU handling in pe_synth_cuda_u64_cones.cu

- Removed redundant thread creation in CUDA evaluation functions to reduce CPU overhead.
- Introduced a work enqueueing mechanism for non-blocking stream execution, allowing concurrent GPU operations.
- Enhanced the `bitset_matrix_device` structure with a host staging buffer to optimize memory management.
- Updated resource cleanup logic to ensure proper handling of CUDA device states and buffers.

Author: MacroModel