Refactor `matrix_ops` public functions as backend-agnostic wrappers

[jc-macdonald]

Summary

After extracting ScipyLinearBackend, replace the extracted implementation in matrix_ops.py with thin delegation to a module-default backend instance.

Implementation

from .backends import ScipyLinearBackend

_DEFAULT_BACKEND = ScipyLinearBackend()

def implicit_solve(left_op, right_op, x):
    """Solve L @ y = R @ x using the default backend."""
    return _DEFAULT_BACKEND.solve(left_op, right_op, x)

def clear_implicit_solver_cache():
    """Clear the default backend solver cache."""
    _DEFAULT_BACKEND.clear_cache()

What stays unchanged

• Operator builders (build_laplacian_tridiag, build_implicit_euler_operators, etc.)
• Stage operator factories (make_stage_operator_factory, make_constant_base_builder)
• Grouping utilities, smooth(), kron_*
• Operator type alias

Files to modify

• src/op_engine/matrix_ops.py — replace extracted implementation with delegation

Acceptance criteria

• from op_engine import implicit_solve, clear_implicit_solver_cache still works
• Existing downstream code (flepimop2 adapter, examples) unaffected
• All tests pass

Dependencies

Depends on: #20 (ScipyLinearBackend)
Parallel with: #21 (inject into CoreSolver)