spec 1724: Builtins/types panel re-audit to ground-truth green

[tamnd]

Re-ran the whole Builtins/types panel (spec 1724) under the 1726 bridge to get
ground truth, then fixed everything that turned out to be a real gopy bug. The
panel is now 28 of 29 files fully green; the one file with residuals (test_dict)
is down to two documented implementation-detail divergences.

What was actually broken

• int_max_str_digits leaked across subinterpreters. The limit lived in a
  package global, so a child interpreter from _testcapi.run_in_subinterp could
  change it and leave the parent clobbered. CPython keeps it on the interpreter
  state. Added a snapshot/restore hook driven from PushSubinterp/PopSubinterp and
  registered it from module/sys. Fixes
  test_int.test_int_max_str_digits_is_per_interpreter.

• Type creation ordered the synthesized dict/weakref descriptors ahead
  of the class body. A class that declares its own __dict__ lost to the
  generated one. CPython installs those getsets after copying the namespace in
  and uses SetDefaultRef semantics so a user slot wins. Deferred the install
  behind a name guard. Fixes test_builtin.test_namespace_order.

• The default breakpoint hook left a stale thread exception. When
  PYTHONBREAKPOINT names an unimportable target the hook swallows the
  ModuleNotFoundError in favour of a RuntimeWarning, but did not clear the thread
  exception the way sys_breakpointhook calls PyErr_Clear, so a later handler saw
  the stale error. Fixes test_builtin.test_envar_unimportable.

• Two test-only C helpers were missing. getbuffer_with_null_view
  (test_bytes) and dict_getitem_knownhash (test_dict CAPI test). Ported both,
  plus a KnownHash lookup that raises KeyError on a miss.

Left as documented residuals

test_dict keeps two failures, both @cpython_only-style implementation details:

• test_splittable_popitem inspects the PEP 412 split-table dict layout through
  sys.getsizeof. gopy uses a single combined representation, so there is no
  split-to-combined growth to observe.
• test_oob_indexing_dictiter_iternextitem relies on the items iterator
  recycling one result tuple and decreffing the previous value to zero mid-loop,
  firing __del__ synchronously. gopy's IterNext returns a borrowed reference
  (spec 1727) and allocates a fresh tuple per advance; the in-place reuse that
  would match CPython previously broke sorted(d.items()).

Both would mean reversing a deliberate gopy design choice for no Python-visible
behaviour, so they stay as residuals with citations in the spec rather than a
shim. The full reasoning is in website/docs/specs/1700/1724.

Ground truth for the whole panel (run from test/cpython/): every file OK except
the two test_dict residuals above.

[tamnd]

CI is green across the matrix (macOS, Linux, Windows) plus vet, lint and cfg-phase-parity.

For the record, the full Builtins/types panel run from test/cpython/ this build:

• 28 of 29 files OK (test_buffer stays out of scope, C buffer-struct API)
• test_dict: 118/120, the two remaining are the documented split-table and iterator-del-timing residuals

The two residuals are impl-detail only and chasing either would reverse a deliberate design choice (no PEP 412 split tables; borrow-model iterator from spec 1727), so I left them documented with citations rather than shimmed.