Fix globalcontext_destroy leaks

[petermm]

Surfaced when doing test_main.c work for zephyr CI.

Knew they were there, should help the esp32 test_main.c harness, especially on the low memory boards and wifi tests.

Mostly (only?) relevant in those scenarios.

note that there is still a leak for: "atom_table_destroy still does not free atom strings allocated with AtomTableCopyAtom."
llm tells me:

I’d estimate 30-60 minutes for a clean fix plus focused tests. The likely shape:
Add an ownership bit to the internal atom-table node, e.g. bool owns_key.
Set owns_key = true only when AtomTableCopyAtom allocates buf.
Set owns_key = false for BEAM/default atoms that point into other owned storage.
In atom_table_destroy, walk all node groups and free((void *) node->key) only for owned keys.
Add/update test-structs coverage for copied atoms.
Let CI/valgrind prove the leak is gone.
Risk level: low to medium. The main thing is not accidentally freeing atom strings that belong to module binaries/default atom storage. But since HNode is private to atom_table.c, the API impact should be tiny.

but that is punted for a follow up.

These changes are made under both the "Apache 2.0" and the "GNU Lesser General
Public License 2.1 or later" license terms (dual license).

SPDX-License-Identifier: Apache-2.0 OR LGPL-2.1-or-later

[pguyot]

We probably can squash this in a single commit