Add golden-output regression tests for mPIE and qNewton

[ShantanuKodgirwar]

What this does

Adds golden-output regression tests that pin the numerical output of mPIE and qNewton.

There is currently no test that runs a reconstruction and checks its result. That makes any
optimisation of the engines unverifiable — you cannot tell a speedup from a silent change in
what the library computes. This PR is the safety net that later performance work is checked
against.

No PtyLab/ source is touched. Tests only.

How it works

Each config runs a few iterations on a deterministic synthetic dataset built in-test, then
compares object, probe and error against a recorded .npz baseline.

The dataset is synthesised rather than loaded from example_data/, so the baselines are
reproducible in CI.

Configs covered

name	engine	modes (nlambda, nosm, npsm, nslice)
mpie_single	mPIE	1, 1, 1, 1
mpie_mixed_state	mPIE	1, 2, 3, 1
qnewton_single	qNewton	1, 1, 1, 1

The mixed-state config is the important one: it exercises the 6D mode broadcasting that a
single-mode-only suite would let a broken change through.

Pinning the randomness

Three sources of nondeterminism, all pinned:

• params.positionOrder = "sequential" — "random" shuffles via the global RNG.
• np.random.seed() before initializeObjectProbe() — it adds noise to break mode degeneracy.
• np.random.seed() again before reconstruct() — mPIE fires its momentum update on
  np.random.rand().

Verified that the baselines reproduce bit-exactly, and that a 0.1% change to betaObject
is caught.

GPU-vs-CPU test

Asserts the two backends agree, using a Frobenius-norm relative error rather than an
elementwise one.

Elementwise is the wrong metric here: these arrays contain near-zero elements where a 1e-7
absolute wobble reads as a huge relative error. Measured divergence across these configs is
1e-8 to 2e-4, so the 1e-3 bound leaves roughly 5x headroom while still catching a
genuinely wrong kernel.

Why ePIE and e3PIE are absent

Deliberate — both would record a meaningless baseline today:

• ePIE.reconstruct() is a generator that callers do not iterate, so it returns without
  doing any work.
• e3PIE raises on its own betaProbe.

They are added in the follow-up PRs that repair them.

Re-recording

After an intended numerical change:

PTYLAB_REGEN_GOLDENS=1 uv run pytest tests/regression -q

Then re-run to confirm the new baselines reproduce, and review the .npz diff before
committing.