Package libcuopt_client separately so the gRPC client installs without CUDA

Description

#1804 makes libcuopt_client.so CUDA-free at the binary level — no cuda, rmm or raft in NEEDED, and no undefined cuopt:: symbols. But it does not yet make a CUDA-free install possible, which was the motivation. Installing the client still pulls the whole CUDA stack, because the packaging has not been split.

Two independent gaps, at two different layers.

1. Conda / wheel packaging

conda/recipes/libcuopt/recipe.yaml has exactly two outputs, libcuopt and libcuopt-tests. libcuopt_client.so ships inside libcuopt, whose run: requirements include:

- cuda-version
- librmm
- cuda-nvrtc
- libcudss

So the library cannot pull CUDA in, but the package it ships in does. This needs a separate libcuopt-client output (and the equivalent wheel) whose run deps are limited to grpc/protobuf/rapids_logger.

2. Python extension modules

This is the larger piece, and it is a separate problem from the .so linkage. The motivating consumer imports only Client, DataModel, Read and SolverSettings. Those are Cython extension modules in the cuopt wheel, and python/cuopt/pyproject.toml declares:

cudf==26.10.*
cupy-cuda13x[ctk]
pylibraft==26.10.*
rmm==26.10.*

Those are Python-level dependencies. They apply regardless of what the extension modules link against, so pointing them at cuopt::cuopt_client alone changes nothing about what pip installs.

Reaching a GPU-free client install needs the host-only extension modules (data_model, solver_settings, io, gRPC client) to link cuopt::cuopt_client and live in a package that does not declare the RAPIDS Python deps. Note that today only routing and distance_engine set linked_libraries explicitly; the LP/settings/io modules would need the same treatment.

Only then can the MCP server depend on that package instead of cuopt.

Sequencing

Worth settling the packaging boundary after #1622 (split libcuopt into cuopt_base / cuopt_routing / cuopt_lp component libs), not before. That PR is renaming cuopt_lp to cuopt_mathematical_optimization and reshaping what the components are, so fixing package names now risks naming them twice. cuopt_client is plausibly a fourth component in that scheme rather than a standalone addition.

Acceptance

• libcuopt-client conda output and wheel, with no CUDA/rmm/cudss run deps
• host-only Python extension modules linking cuopt::cuopt_client, packaged without cudf/cupy/rmm/pylibraft
• a clean-environment check that installs the client package and imports Client, DataModel, Read, SolverSettings with no CUDA runtime present
• MCP server depending on the client package

Relationship to #1635

#1635 tracks splitting the libcuopt wheel into base / routing / mathopt / grpc packages once #1622 merges. cuopt_client is a fifth component in that same split, and the two share almost all of their machinery: new ci/build_wheel_libcuopt_*.sh scripts, new python/libcuopt_*/ package directories, per-package dependency sets in dependencies.yaml, and conda recipe outputs.

These should be done together rather than separately. The distinction worth preserving is that cuopt_client is the only component with no CUDA dependency at all — measured 2.3 MB stripped, needing only grpc/protobuf/abseil — so it is the one that makes a GPU-free install possible, while the others are motivated by the PyPI size limit.

Related: #1801, #1802, #1803, #1804, #1622, #1635.