Generate the routing gRPC field mapping from a single source of truth

[ramakrishnap-nv]

Follow-up to review feedback from @rg20 on #1597 (mapper, to_device, round-trip test).

Problem

Adding one field to cpu_routing_problem_t today means editing five places by hand:

1. the struct — cpp/include/cuopt/routing/cpu_routing_problem.hpp
2. the wire format — cpp/src/grpc/routing/cuopt_routing.proto
3. host → proto — map_routing_problem_to_proto
4. proto → host — map_proto_to_routing_problem
5. host → device — cpu_routing_problem_t::to_device

Nothing enforces that all five stay in sync. Not the compiler, not a test.

The failure mode is worse than the usual boilerplate complaint. A forgotten field
does not crash — the value silently never crosses the wire, and the remote solve
returns a plausible answer for a different problem than the one submitted.
That is expensive to detect and easy to miss in review.

Scale: cpu_routing_problem_t currently has 41 std::vector members, 6
std::map members, and 5 nested struct types. Roughly 80% of that is one
regular pattern repeated 41 times.

Suggested direction

Both review threads share a single root cause — the field list is written out
five times — so it is worth fixing once rather than twice.

Single source of truth for the routing field list, expanded into the proto,
both mapper directions, to_device, and a round-trip test. Either:

• a routing section in cpp/src/grpc/codegen/field_registry.yaml, reusing
  generate_conversions.py (which already emits 27 .inc files and supports
  repeated_messages), or
• an X-macro field-list header, which is less machinery and also covers
  to_device, which is outside the gRPC codegen's remit.

Caveat worth designing around: routing is less regular than LP. The 6 maps carry
real semantics — order_service_times is keyed by vehicle id with -1 meaning
"all vehicles", vehicle_breaks maps to a vector of structs. Expect ~80%
generated plus a tail of hand-written special cases, not a pure win.

Cheaper first step

A round-trip property test — populate every field with distinct values, go
host → proto → host, compare — catches the same class of bug for a fraction of
the effort and does not restructure anything. Worth doing first, and worth
keeping even after codegen lands.

Prior art in-tree: python/cuopt/cuopt/tests/routing/test_routing_grpc_serialization.py
already fails loudly when a new cuopt.routing._deferred._SETTERS entry is
unmapped. That guards the Python DataModel → cpu_routing_problem_t direction;
this issue is about the remaining C++ hops.

Not blocking #1597

#1597 is approved with these threads open, so this is design follow-up rather
than a merge blocker.