loop-induced output: fix standalone, refuse the formats with no MadLoop backend

[oliviermattelaer]

output standalone crashed for any loop-induced ([noborn=…]) process. Two
separate bugs, plus a decision that has been revisited now that the first one is
understood.

The reported crash was not the real bug

The symptom was an IndexError in write_check_sa's flavour-grouping block, which
looked like a flavour-grouping problem. It was not: legs_with_decays was correct all
along (4 entries for g g > z z); the flavour tuples had 6.

_flavor_enumeration_context (madgraph/core/helas_objects.py) collects every
motherless wavefunction, and in a LoopHelasMatrixElement the L-cut wavefunctions are
motherless too, carrying number_external = nexternal+1/+2. Instrumented on base:
pdgs=[21, 21, 23, 23, 1, -1]. LoopHelasMatrixElement.get_nexternal_ninitial already
carries exactly the not wf.get('is_loop') filter for this reason; the flavour
enumeration never got it.

Fixing that exposes the second bug: ExportV4Factory's format.startswith('standalone')
branch returned the tree-level ProcessExporterFortranSA unconditionally, unlike the
two madevent branches below it. The output then died in the tree-level
write_matrix_element_v4 with "wavefunction_rank has not been computed".

Allowed and refused

A previously-shelved decision (refuse loop-induced output everywhere it has no backend)
is folded in here, minus its standalone entry — because standalone can now work.

Allowed: madevent, standalone, plugin.
Refused with an InvalidCmd naming [sqrvirt=]: matrix, standalone_msP,
standalone_msF, standalone_rw, standalone_cpp, matchbox, matchbox_cpp, mg7,
mg7_v5, standalone_mg7. aloha is unaffected — it returns before the check.

standalone is a prefix shared by formats that do not work, so every membership test
is exact (format not in LOOP_INDUCED_FORMATS), never startswith. The one surviving
startswith('standalone') — the factory branch selector — has the exact test inside it.
test_loop_induced_formats_are_matched_exactly fails if a future refactor "simplifies"
that back to a prefix.

A refusal must never destroy data, so the check also sits in do_output above the
shutil.rmtree that wipes an existing output directory; the factories run after it. That
covers --me_exporter= too, which writes into the same directory and would otherwise
reach only the factory backstop — verified by test that the directory survives both ways.

validate_model

generate g g > h [noborn=QCD] as the very first command died with
AttributeError: 'NoneType' object has no attribute 'merged_particles'. A guard on that
line alone would have moved the crash three lines down — the rest of the function
dereferences _curr_model five times. The real defect is ordering: master_interface
calls validate_model directly for noborn, before create_loop_induced runs
check_add, whereas [virt=]/[real=] arrive only after check_generate has imported
the default model. Fixed by importing the model at the top of validate_model exactly as
check_generate does; the sm → loop_sm upgrade below still runs.

Validation

• Builds and runs, not merely generates: g g > z z [noborn=QCD] →
  2.5927822006924310E-004 (MadLoop rc 217, stable), bit-identical to the same process
  generated as [sqrvirt=] on base — which reaches the MadLoop interface and was never
  broken. g g > h h → 3.2829343688358318E-005, matching the recorded reference;
  g g > h → 9.3702613225405892E-003.
• Nothing else moved: PYTHONHASHSEED=0, base vs branch, check_sa.f and matrix.f
  byte-identical for g g > t t~ (plain) and p p > e+ e- (merged particles), with the
  merged flavour table keeping its sign transfer (FLAVOR(2,2)=2 → PDG -2). The only
  generated-code change anywhere in a loop-induced madevent tree is a comment,
  C FLAVOR = [1, 1, 1, 1, 1] → C FLAVOR = [1, 1, 1] — the bug becoming visibly fixed.
• Tests: tests/unit_tests/loop/test_loop_induced_output.py, 16 tests. 12 fail on
  base; the 4 that pass are exactly the controls (madevent, sqrvirt, virt, the
  tree exporter). 16/16 on branch. Plus testIO_loop_induced_standalone_output, which
  fails on base with the original IndexError.

One pre-existing bug found, worked around rather than fixed

Building a loop matrix element with compute_loop_nc=False poisons a process-global
colour memo that a later loop-induced output madevent reuses, dying in
get_icolamp_lines on v[4]-v[5]. Root cause: ColorBasis._canonical_dict
(madgraph/core/color_amp.py:51) is a class attribute; ColorBasis.__init__ shadows
it with an instance dict but LoopColorBasis.__init__ never calls its parent's
__init__, and the cached loop_Nc_power is neither part of the cache key nor
recomputed on a hit. Reproducible on clean main with a plain user script, and this
branch does not make it worse — the route opened here sets compute_loop_nc=True and so
cannot poison the cache. The tests build with compute_loop_nc=True to stay clear of it,
matching what group_subprocs does in production.