refactor[next]: drop factory-boy for plain builders + sub-component injection

[egparedes]

Summary

factory-boy (a test-data library) was repurposed in production to compose the GTFN and DaCe backends and their OTF compilation workflows. Since every model is already a plain frozen dataclass, the Trait/SubFactory/SelfAttribute/LazyAttribute machinery and the stringly-typed __-path overrides added indirection without value.

This PR (two commits) removes that and replaces it with plain builders + dependency injection:

1. Drop factory-boy → plain builder functions / direct dataclass construction; remove the runtime dependency.
2. Inject sub-components instead of forwarding nested config through maker kwargs.

It implements two items from the layered-architecture refactoring proposal: "plain-code backend builders, not factory-boy" and decoupling builders from their sub-components' internals.

Commit 1 — remove factory-boy

• Leaf steps → direct construction: removed CompilerFactory, GTFNTranslationStepFactory, DaCeTranslationStepFactory, DaCeCompilationStepFactory.
• GTFN/DaCe: factory classes → make_gtfn_workflow / make_gtfn_backend / make_dace_workflow / make_dace_backend. Traits became bool kwargs; __-path overrides became real parameters.
• factory-boy moved from runtime dependencies to the test group (still used by the cartesian/eve IR test-data factories); stale mypy override removed; uv.lock refreshed.

Commit 2 — sub-component injection (lightweight DI)

Flattening the __-paths into explicit maker kwargs re-coupled each maker to its children (most visibly make_dace_backend, forwarding ~6 translator-only knobs). Commit 2 fixes that: a maker exposes only cross-cutting configuration (device, caching, auto-optimize, build type); a single-component knob is set by injecting that component.

# before:  make_gtfn_backend(enable_itir_transforms=False)
make_gtfn_backend(translation=GTFNTranslationStep(enable_itir_transforms=False))

• workflow.with_changes(step, **changes) — guarded helper that stamps cross-cutting fields (device_type; dace auto_optimize) onto a step only when it is a dataclass declaring them, so injecting an arbitrary callable stays safe.
• GTFN makers drop enable_itir_transforms/use_imperative_backend, gain translation/bindings/compilation/decoration injection params.
• DaCeTranslator gains defaults for its formerly-required fields; a new make_dace_translator(...) owns the translator-local knobs and the relocated optimization_args validation / unit_strides_kind derivation.
• make_dace_workflow/make_dace_backend shed their translator-forward kwargs and accept injection.

Behavior & compatibility

• All pre-built backends keep identical names and behavior (run_gtfn_*, run_dace_*, and the gtfn_cpu/gtfn_gpu re-exports) — verified field-by-field.
• Breaking change: make_dace_backend no longer accepts the translator-local kwargs (async_sdfg_call, optimization_args, use_metrics, use_zero_origin, unstructured_horizontal_has_unit_stride, use_max_domain_range_on_unstructured_shift). External callers (e.g. icon4py) switch to make_dace_backend(..., translation=make_dace_translator(...)).

Verification

• uv run mypy src/ clean; full pre-commit (ruff, ruff-format, tach, mypy, license, uv-lock) green on all changed files.
• GTFN + otf unit suites pass (incl. a new test_backend_inject_translation); the migrated DaCe CPU tests (test_dace_backend, test_dace_bindings) pass; make_dace_translator's validation paths checked directly.
• The GPU parametrizations (exec_alloc_descriptor1, test_dace_fastcall) can't run locally (no cupy/GPU) — needs a GPU runner for full-matrix confidence.

Note on typing

Constructing OTFCompileWorkflow directly surfaces two CachedStep.persistent(...) calls where mypy cannot solve the HashT type variable (it appears only in the fingerprinter's return) — previously hidden by factory-boy's dynamic typing. These carry a scoped, documented # type: ignore[arg-type].

The Backend → toolchain rename and the pipeline collapse from the larger refactor are intentionally left for follow-ups.