infer_rel_schema emits both inputs plus the mark column for mark joins

[alexandrefimov]

_join_column_shape sends both mark join types to both+mark, and the output is then built as left input, right input, mark column (type_inference.py:558 and :575; same on main). Join Operation (v0.102.0), Direct Output Order: "For mark joins, the emit order is either left or right with a "mark" column appended at the end." So LEFT_MARK emits the left input plus the mark column, RIGHT_MARK the right input plus it, and the opposite input is in neither.

Reproduced with 0.31.0 through the builders (mark below is the library's own JOIN_MARK_COLUMN_NAME):

import substrait.algebra_pb2 as stalg, substrait.type_pb2 as stt
from substrait.builders.plan import read_named_table, join
from substrait.builders.extended_expression import literal
from substrait.extension_registry import ExtensionRegistry
from substrait.type_inference import infer_plan_schema

REQ = stt.Type.NULLABILITY_REQUIRED
reg, i64 = ExtensionRegistry(), stt.Type(i64=stt.Type.I64(nullability=REQ))
ns = lambda *n: stt.NamedStruct(names=list(n), struct=stt.Type.Struct(
    types=[i64] * len(n), nullability=REQ))

p = join(read_named_table("l", ns("l0", "l1")), read_named_table("r", ns("r0", "r1")),
         literal(True, stt.Type(bool=stt.Type.Boolean(nullability=REQ))),
         stalg.JoinRel.JOIN_TYPE_LEFT_MARK)(reg)

print(list(infer_plan_schema(p, registry=reg).names))
# actual:   ['l0', 'l1', 'r0', 'r1', 'mark']
# expected: ['l0', 'l1', 'mark']

substrait-java derives left-plus-mark for that same plan and refuses to load it, which is how I ran into this: Plan.Root names count (5) must match input record type depth-first named-field count (3).

One thing I did not want to assert: the join condition still references fields of both inputs, so narrowing the output should leave condition-side numbering alone, but that is worth a second pair of eyes. Happy to send the PR if you want it.