Struct-valued expressions other than column() emit a single output name #312

ExpressionReference.output_names holds field names in depth-first order, so a struct-valued expression needs one name for itself plus one per nested field. Apart from column(), the expression builders in src/substrait/builders/extended_expression.py always produce a single name, whatever the output type is. A projected struct-valued function, cast, literal or if_then therefore yields a RelRoot.names / inferred NamedStruct with too few names, and nothing raises an error:

from substrait.builders.extended_expression import cast, column
from substrait.builders.plan import read_named_table, select
from substrait.builders.type import i64, named_struct, struct
from substrait.extension_registry import ExtensionRegistry
from substrait.type_inference import infer_plan_schema

registry = ExtensionRegistry(load_default_extensions=False)
ns = named_struct(
    names=["a", "s", "x", "y"],
    struct=struct([i64(), struct([i64(), i64()])], nullable=False),
)
plan = select(
    read_named_table("t", ns), [cast(column("s"), struct([i64(), i64()]))]
)(registry)

list(plan.relations[-1].root.names)                  # ['cast(s)']  needs 3 names
list(infer_plan_schema(plan, registry=registry).names)  # ['cast(s)']  for a struct<i64, i64>

Fixing this means deriving each builder's output type and padding output_names to type_num_names of it. The spec does not say what the nested fields of a computed struct should be called, so before picking a scheme, check what the sibling bindings (substrait-java, substrait-rs) generate and state the choice in the PR. alias() renames only output_names[0], which is correct once the nested names are present.

Follow-up to #311, which fixed select, project and aggregate to carry a struct column's full name list.