🐛 Engine fixes for Map-zone cloning, gather races, and per-task (de)serializers

[elinscott]

Summary

Three engine-level fixes uncovered by porting a production workflow with per-orbital Map zones. Each failure mode below is independent; they ship together because the reproducing workflow unearths all three.

1. @task decorator: spec-level deserializers / serializers

@task now accepts deserializers= / serializers= kwargs and stores them on the spec. PyFunctionTask.execute falls back to the spec-level values when no per-call kwarg is provided, in both the sync and async execution branches.

Motivation: provides a mechanism to opt out of the default StructureData -> Atoms auto-deserializer

@task(deserializers={"aiida.orm.StructureData": "my_plugin.utils.identity"})
def analyze(structure):
    ...

Only meaningful for plain-Python (PyFunction) tasks; CalcJob / WorkChain / calcfunction tasks are unaffected (AiiDA owns their deserialization).

2. _patch_cloned_tasks: materialise dynamic-namespace sockets on cloned tasks

Map-zone task cloning looked up link targets with a bare to_node.inputs[scoped_name]. A dotted path into a dynamic namespace (e.g. pseudos.O) does not exist on a fresh clone, so relinking raised AttributeError at submit time. Cloning now resolves the target through Graph._resolve_or_create_input_socket, which materialises the child socket on demand.

3. update_map_task_state: gather race and failure propagation DROPPED (see comment below)

Two related failure modes when a Map zone's iterations finish close together:

• Race: are_childen_finished can return True for the map zone before every per-prefix gather_item clone has recorded its result in ctx._task_results. The gather loop then KeyErrors. The method now bails out cleanly while any clone is still in flight; the next continue_workgraph tick re-enters once results are populated.
• Failure masking: a failed mapped iteration leaves its gather_item clone SKIPPED (on_task_failed skips downstream children), so a check keyed on FAILED alone never fires and the gather loop KeyErrors on the empty result, burying the original calculation failure under an unrelated trace. Gather clones that end SKIPPED or FAILED now propagate FAILED to the map task, preserving the real cause. Covered by a new test, test_map_zone_failed_iteration: KeyError before the fix, map task ends FAILED after.

Compatibility

Fix 2 requires Graph._resolve_or_create_input_socket, which is not yet in a released node-graph (introduced alongside the dynamic-namespace work targeting node-graph main). Draft until that release; fixes 1 and 3 have no new dependencies and can be split out if preferred.

[elinscott]

Dropped fix for the race condition from this PR as it is separately addressed by #776