Fix backend crash in age_create_barbell_graph with a null node label

[NotHimmel]

Backport of a fix for the 1.7.0 line. Upstream: apache#2521 (PR), apache#2519 (issue).

Problem

node_label is declared name = NULL in the SQL signature of
ag_catalog.age_create_barbell_graph(), so leaving it out — or passing NULL
explicitly — is a supported call. Both crash the backend with SIGSEGV, which
makes the postmaster reinitialize and drops every other session on the
instance.

SELECT create_graph('barbell_test');
SELECT age_create_barbell_graph('barbell_test', 5, 0);
-- server closed the connection unexpectedly

LOG: client backend (PID 312566) was terminated by signal 11: Segmentation fault
LOG: all server processes terminated; reinitializing

Root cause

Two separate faults on the null-node_label path.

1. The default label was copied into a null pointer.

Namenode_label_name=NULL;
...
if (PG_ARGISNULL(3))
namestrcpy(node_label_name, AG_DEFAULT_LABEL_VERTEX);

namestrcpy() writes through the pointer it is given and does not allocate.
The fix gives the default its own NameData, which also makes the default
actually take effect — it never did.

The immediate segfault only happens on PostgreSQL 14 and later. namestrcpy()
used to start with if (!name || !str) return -1;, removed by PostgreSQL
commit 1784f278a638, first released in 14. On PG 13 and earlier the call
silently did nothing and the failure moved downstream.

2. The node label was forwarded as the raw argument Datum.

DirectFunctionCall4(create_complete_graph, ..., arguments->args[3].value);

DirectFunctionCall4() marks every argument as not null, so a null node label
reached create_complete_graph() as a non-null NULL pointer and was
dereferenced by its vertex/edge label comparison. Fixing only the
namestrcpy() call is not enough — verified: with just that change,
age_create_barbell_graph('g', 5, 0, NULL, NULL, 'E') still segfaults.

create_complete_graph() already handles its own null node label correctly;
this change makes the barbell function follow the same approach.

Introduced by the original barbell implementation, 0c79370
("Barbell graph generation", apache#648, 2023-02-17).

Testing

Extended regress/sql/graph_generation.sql with the previously untested cases.
The existing barbell tests always passed a node label, except for the
all-arguments-null case, which errors out on the graph name before reaching
this code — which is why this was never caught.

On PostgreSQL 18.4, built from source:

• Without the code change, graph_generation fails with
  server closed the connection unexpectedly, and because the crash takes the
  temporary instance down with it, the tests that follow fail as well.
• With the code change, # All 31 tests passed.

The new cases also confirm the default label is applied: gp7 gets 10 vertices
under _ag_label_vertex (two K5s) and 21 edges (2 × 10 plus the bridge), and
the three-argument call now returns ERROR: edge label can not be NULL
instead of crashing.

Out of scope

Three other defects in the same function, left alone to keep this focused:
if (PG_ARGISNULL(1) && PG_GETARG_INT32(1) < 3) should use || (a
graph_size of 1 or 2 passes silently); node_properties and
edge_properties are accepted but ignored; bridge_size is validated but
unused, as the in-code comment notes.