loop-induced: allow a polarised massive particle, and cover the last-leg frame

[oliviermattelaer]

Stacked on #79. Lifts the refusal of a polarised massive particle for the
loop-induced NLO modes ([noborn=…], [sqrvirt=…]).

Why the restriction was an omission

Loop-induced processes are exported through the LO madevent template, so they
already have me_frame and the LO boost — nothing needed building. Verified in the
generated code rather than argued: auto_dsig1.f carries
IF(FRAME_ID.NE.6) CALL BOOST_TO_FRAME(PP, FRAME_ID, P1) and passes P1 to
SMATRIX1, which hands it to ML5_0_0_1_SLOOPMATRIX_THRES. Source/run_card.inc
reads FRAME_ID = 8 for me_frame=[3] and 16 for [4].

Measurement

g g > z{0} z{0} [noborn=QCD], shipped card, nevents=100:

me_frame	FRAME_ID	sigma
[3]	8	2.543e-02 +- 9.901e-05 pb
[4]	16	2.543e-02 +- 9.901e-05 pb

results.dat byte-identical between the two, which is the right expectation:
the two Z bosons are identical and both longitudinal, so the frames must agree
pointwise. Before PR #91 this pair read 2.543e-02 vs 3.211e-02 — that PR fixed
improve_ps dumping the momentum-conservation residual on leg NEXTERNAL and
destroying the exact zero the polarisation axis depends on. Both numbers here were
reproduced independently by a second agent on a fresh tree.

Tests

• test_polarised_loop_induced_me_frame (me_frame=[3]) and
  test_polarised_loop_induced_me_frame_last_leg ([4]), both wired into
  acceptancetest_107 by explicit name — the workflow invokes tests by name, so an
  unwired test never runs. The [4] test was added skipped while PR MadLoop: keep an external leg exactly at rest through improve_ps #91 was
  outstanding; the skip is deleted here.
• The assertions are not weakened. check_parton_output demands
  |Δ|/√(5e-4² + σ²) < 3; the measured 0.02543 sits 0.18 sigma from the asserted
  0.02534, an unboosted run (5.805e-02) would be ~47 sigma out, and a PR MadLoop: keep an external leg exactly at rest through improve_ps #91 revert
  (3.211e-02) ~9-13 sigma out.
• .github/workflows/unittest.yml now also runs test_polarisation_nlo_regimes,
  the base branch's dedicated test for the three-regime split, which was defined but
  never invoked by CI.

One test assertion changed direction, deliberately

g{L} g > z z [noborn=QCD] moved from the rejected list to the accepted one. That is
a consequence of a base-branch decision, not a new one here: base commit
ee63ba492 scoped the colour refusal to the subtracted regime and added its own
positive assertion (g g > t{L} t~ [noborn=QCD]) with the comment "colour is no
restriction where there is no subtraction". The line added here is redundant
coverage of an assertion the base already makes. Upstream main has no colour check
at all; the refusal is this chain's own invention, narrowed to where the FKS-emitter
hazard actually exists.

The doc previously said the colour refusal "applies in every mode and stays", which
has been false since the rebase and is corrected — that sentence is exactly what a
future conflict resolver would read to justify restoring a blanket colour check,
which would silently re-close this feature.

Stated limitation

No coloured polarised loop-induced process has been run anywhere in this chain.
Nothing in the LO boost path assumes a colourless polarised leg — boost_to_frame is
purely kinematic and colour enters nowhere between the boost and the matrix element —
and the specific case added here (g{L}, massless) needs no frame at all, since
massless helicity is Lorentz invariant. But permitting it is a format decision, not a
validated one, and the doc now says so.

Notes for the merger

• Commit 05a7887b1 is nearly empty: the guard it introduced went upstream into the
  base, so only the test_cmd.py additions survive in it, and its message still
  describes the guard change. History was left intact rather than rewritten; harmless
  under a squash merge.
• acceptancetest_107 runs both acceptance tests in one job — 23.4 min + 21.0 min on
  an 18-core machine, so likely longer on a CI runner. Worth splitting if that slot
  becomes a bottleneck.