refactor[next]: replace factory-boy factories with plain builders, fixing run_gtfn_imperative

[egparedes]

factory-boy is a test-data library, but gt4py.next used it in production
to compose the GTFN and DaCe backends and their compile workflows. Every
object it built is already a frozen dataclass, so the
Trait/SubFactory/SelfAttribute/LazyAttribute machinery added a second
construction language that no type checker can see.

Replace the seven factory classes with plain builder functions and move
factory-boy to the test dependency group (the cartesian and eve IR
test-data factories keep using it for its intended purpose).

Builders follow two rules: a builder takes cross-cutting configuration only
(device, caching, build type, auto-optimize) and configures the steps it
creates from it; an injected sub-component is used verbatim. The second rule
needs a guard, so workflow.check_device_agreement raises when an injected
step's device disagrees with the pipeline being built. It inspects only steps
that structurally declare a device (workflow.DeviceConfigurable) and is used
to check, never to mutate.

Behavior change: run_gtfn_imperative becomes imperative for the first time.
It was declared with otf_workflow__translation__use_imperative_backend=True,
but translation is a LazyAttribute rather than a SubFactory, so
factory-boy dropped the deep path silently. (The sibling override on
run_gtfn_no_transforms did take effect precisely because bare_translation
is a SubFactory — that asymmetry is the whole bug.) The backend was a
duplicate of run_gtfn under another name, so the GTFN_CPU_IMPERATIVE
test-matrix entry had never exercised imperative code generation.

Switching it on exposes a pre-existing IR defect: the imperative code path
leaves CSE temporaries undeclared, so symbol validation rejects the IR. That
is tracked as issue #2810 and is out of scope for a construction refactor, so
the two call sites that hit it -- test_hdiff and
test_concat_where::test_lap_like[static_domains] -- are xfailed against it.
test_lap_like[dynamic_domains] does not fold enough to trigger the defect
and is deliberately left running.

Second latent bug: run_gtfn_no_transforms.name was run_gtfn_cpu,
colliding with run_gtfn. It is now run_gtfn_cpu_no_transforms. This
rotates no cache -- the build cache keys on the entry-point name plus a
fingerprint of the ExtensionSource, and the translation-cache directory is
keyed on the literal backend family (gtfn / dace). Backend.name reaches
only the metrics source key and one error message, so the collision's real
cost was two distinct backends sharing one metrics identity.

All other pre-built backends are unchanged, verified field-by-field against
the previous construction.

Removing the factories also removed the 8 # type: ignore[assignment] # factory-boy typing not precise enough suppressions in src/, which had been
masking real typing problems. Three are fixed here (device-type narrowing in
both backend builders, and CachedStep.persistent's unsolved HashT at two
call sites). One remains as a scoped, documented type: ignore:
OTFCompileWorkflow is not parameterized over the code spec, so its
bindings field is typed for ProgramSource[Any] while ExtensionGenerator
accepts only C++-like specs. Parameterizing the pipeline is the real fix and
belongs with the pipeline rework.

make_dace_backend deliberately keeps its translator-local keyword
arguments, so external callers are unaffected.

See ADR 0028.