Make U honour its documented matrix and define the phase-gate family exactly

[ciaranra]

Stacked on #615 (which is stacked on #601); review those first.

Summary

The ArbitraryRotationGateable::u doc comment states the standard OpenQASM matrix, but the default implementation was rz(lambda).ry(theta).rz(phi), which differs from that matrix by exp(-i(phi+lambda)/2). The universal gate therefore executed phase-wrong on every amplitude-carrying simulator that inherited the default. Separately, both QASM includes lowered p, phase, and u1 to rz, which is wrong by exp(i*lambda/2): the standard definition is u1(lambda) = U(0,0,lambda) = diag(1, exp(i*lambda)).

This PR makes U honour its own documented matrix and defines the phase-gate family in terms of it, giving the exact identities

p(0) = I,  p(pi/4) = T,  p(pi/2) = SZ,  p(pi) = Z

with RZ(theta) = exp(-i*theta/2) p(theta) documented as the bridge. The rotation family RX/RY/RZ is unchanged.

Changes

• Default U now implements exp(i(phi+lambda)/2) RZ(phi) RY(theta) RZ(lambda), with the scalar carried through amplitude representations. The four state-vector simulators with direct U implementations were verified phase-exact and left unchanged; StateVecSoA32, the GPU state vectors, CuStateVec, StabVec, StabMps, and Mast received exact handling.
• p, phase, and u1 lower to U(0,0,lambda) in both includes. cp, cu1, and cphase were verified column-by-column as exactly diag(1,1,1,exp(i*lambda)) — controlled constructions are where these phases become observable, so this was checked rather than assumed.
• R1XY was audited for the same defect and does not have it: its outer RZ angles sum to zero, so the composition equals its documented matrix exactly. A column-exact test now locks that in.
• Exact testing exposed two stale-cache defects in StabVec, both fixed: CliffordFrame::is_diagonal accepted frames that send Z to -Z (X and Y counted as diagonal), and reset did not clear frame state.
• Phase-blind representations (tableau, density-matrix, propagation, compile-only) are unchanged, as is the rotation family.

Mutation evidence

Restoring the old rz.ry.rz default fails all four U guards: U(0,0,pi/4) carries exp(-i*pi/8), U(0,0,pi/2) carries exp(-i*pi/4), U(0,0,pi) carries exp(-i*pi/2), and the zero-angle case is caught by a scalar-hook witness (call count 0 against expected 1), since the matrices coincide at zero. Restoring p to rz fails all four QASM guards with a structural witness observing RZ instead of U.

Verification

• Full debug and release suites for pecos-simulators, pecos-qasm, pecos-stab-tn; clippy --all-targets -D warnings on all five touched crates; cargo fmt --check; git diff --check.
• Python suite: 4,677 passed. The 50 failures are all Selene TCP-stream, multiprocessing, and network-dependent doc-build tests denied sockets by the sandbox; the failure list was inspected and none touch simulator or phase paths.
• GPU/CUDA suites compile in both profiles; runtime failures are the pre-existing no-hardware results (llvmpipe rejected, CUDA error 100), unchanged in count.
• No trajectory, golden, or snapshot hash changed.