feat: add tiled exact search primitives and optional GPU scoring support

[AxelRolov]

Summary

• add lazy package exports for scoring and optimization to avoid eager import cycles
• add scoring_tiled.py with tiled exact-search utilities for squared norms, batched Tanimoto, and streamed top-k selection
• add an optional CuPy-based GPU kernel bridge plus a gpu extra so CPU-only installs continue to work unchanged

Details

This PR prepares cdr_bench to act as a reusable compute backend for large-scale exact search workloads.

The CPU layer adds:

• compute_fp_squared_norms
• tanimoto_similarity_matrix_batch
• topk_1d / topk_matrix
• exact_tanimoto_topk_from_block
• streaming_exact_topk
• resolve_tanimoto_backend

The GPU layer adds:

• gpu_kernels.py with a safe CuPy import wrapper
• CuPyKernelBridge with a RawModule fused top-k kernel path and a CuPy fallback path
• optional dependency wiring under [project.optional-dependencies].gpu

Validation

• compiled scoring_tiled.py and gpu_kernels.py
• import smoke for from cdr_bench.scoring import compute_fp_squared_norms, streaming_exact_topk
• verified the GPU guard path on a CPU environment returns has_gpu_support() == False instead of breaking imports

Notes

The GPU bridge is intentionally optional and isolated. A CPU-only uv sync should continue to work without cupy installed, while GPU users can opt in with uv sync --extra gpu.