feat: implement Circuit::optimize (basic gate-level optimisation)

[manishpaulish]

Implements Circuit::optimize, which until now was a stub that
panicked. As suggested in #157, this is the lightweight "middle ground"
rather than a full port of pyzx's basic_optimization: it does the
obvious cancellations and commutations and leaves heavier rewriting to
the ZX level.

Approach

• Repeatedly scan the gate list; combine each gate with the first later
  gate it can cancel or fuse with, provided every gate in between
  commutes with it. Run to a fixpoint.
• Cancellation: a gate adjacent to its adjoint (covers self-inverse
  gates and the S/Sdg, T/Tdg pairs).
• Fusion: adjacent same-axis phase gates (ZPhase, XPhase) add their
  phases, dropping to identity when the sum is zero.
• Commutation is conservative: disjoint qubits, or two diagonal gates.
  Measurement, reset, ancilla-init and post-selection are hard barriers.
  This guarantees the pass never reorders non-commuting gates.

Correctness

• Unit tests for each cancellation/fusion/commutation/barrier case.
• A tensor-equivalence test over random Clifford+T circuits, comparing
  the linear maps before and after via equal_circuit_tensor.
• Locally I also verified equivalence on the tensor-checkable circuits
  in circuits/small (tof, barenco_tof, mod5_4, qft, hwb6, ...), all
  equivalent.

Results

• On circuits/small, total gate count drops from 8625 to 2665 (about
  69%). This is total gate count, not T-count; the bulk is Clifford
  cancellation from compute/uncompute structure, which is exactly what a
  basic cleanup pass should catch before ZX simplification.

Possible follow-ups (kept out to stay minimal and correct)

• Fuse S/T/Z aliases by normalising to ZPhase first.
• Exploit CNOT control-line commutation with diagonal gates.