refactor[next]: add stage observability and Toolchain.translate, fix dace __sdfg__ reach-in

[egparedes]

Give the toolchain a sanctioned way to observe and partially run its
pipelines, and remove the one consumer that had to reach into them:

• otf.workflow.stage_hook(name, artifact): an event hook emitted after
  each step by the two named pipelines (backend.Transforms and
  recipes.OTFCompileWorkflow, via the generic MultiWorkflow /
  NamedStepSequence loops) and by Toolchain.translate. Artifacts are
  passed opaquely and never formatted at the emit site; with nothing
  subscribed the hook body is empty, so emission is an empty callback loop
  (~160 ns).
• GT4PY_DUMP_STAGES=<dir> (config.DUMP_STAGES) plus the
  instrumentation.stage_dump subscriber, which writes every stage
  artifact to <dir>/<program>/<NNN>_<step>.<ext>. Registered at import
  time when the variable is set, so a plain program run dumps its stages
  without any user code; enable() / disable() cover programmatic use.
  Exclusive-create with an index bump keeps concurrent compilation workers
  and same-named programs from overwriting each other.
• Toolchain.translate(definition, compile_time_args): the sanctioned
  partial run (frontend + translation only). It narrows to the standard
  OTFCompileWorkflow shape and raises a clear error on a monolithic
  backend. Per-call step options are deliberately not offered; a caller
  needing a variant step builds a variant pipeline with
  dataclasses.replace.
• The dace __sdfg__ reach-in is gone. It duck-typed through
  backend.backend.translation, unwrapped the CachedStep by hand, and
  used object.__setattr__ to write into the in-memory-cached
  past_to_itir stage and its frozen CompileTimeArgs. It is replaced by
  a dace-owned translate-only toolchain variant built with
  dataclasses.replace, plus Toolchain.translate.

That last change also fixes a real bug: because the mutated stage was the
cached one, a second __sdfg__ call re-transformed an already-transformed
program, and any other consumer of that cache entry saw args whose
offset_provider had been replaced by offset-provider types, losing the
runtime connectivity tables. A regression test covers it.

Removing the reach-in drops the last non-doc consumer of the workflow
mixins' .replace, which unblocks the pipeline simplification.

GT4PY_DUMP_STAGES is new opt-in functionality; nothing else changes
behavior, and no fingerprinted dataclass gains a field, so no cache keys
rotate.