[FEA] Routing validation: keep a structural 'simple error set' in Python, move semantic checks to C++

Is your feature request related to a problem? Please describe.

Routing input validation currently happens in two places: python/cuopt/cuopt/routing/validation.py re-checks things (null values, non-negative values, numeric ranges) that the C++ layer already validates authoritatively (e.g. "Cost matrix input cannot be null.", "Array contains negative or 0 values"). That is duplicated maintenance and can drift.

With store-then-build (PR for routing-cpu-api-phase2), the device model — and thus C++ validation — is built at solve time, which makes it natural to treat C++ as the single source of semantic validation.

Describe the solution you'd like

Draw a clean, non-overlapping boundary:

• Python keeps a small "simple error set" — cheap, host-side, structural/type checks that catch silly mistakes early with clear messages and do not require building the device model:
  • reject Python lists/tuples; container/dtype type checks (C++ cannot — it only sees pointers),
  • matrix is square and dim == num_locations,
  • array/size consistency (order arrays == num_orders, vehicle arrays == fleet_size, pickup/delivery pairs equal length, capacity dims),
  • obvious sanity (num_locations >= 1, index-in-range).
• C++ owns everything semantic/numeric (null, negative/zero, solver-specific bounds, feasibility). Remove the duplicated Python checks (_contains_null, validate_positive*, trim validate_range).

This removes the double-checking while preserving fast feedback on structural mistakes.

Additional context

• Tests that currently assert Python validator messages at set-time (test_dist_mat, test_time_windows, test_range) move their pytest.raises to the Solve() call (as test_invalid_datamodel already does), except the structural ones which stay eager.
• Before deleting a Python check, confirm C++ raises an equally clear message; otherwise improve the C++ message rather than lose it.