refactor(corekit): normalise sentinel object names to SCREAMING_SNAKE

[JarryShaw]

The four sentinel objects in pcapkit/corekit/sentinels.py carry three different casings, while all four types already follow CamelCase + Type. Normalise the objects to SCREAMING_SNAKE.

The owner's ruling, verbatim (from #719):

take SCREAMING_SNAKE and accept the breaking change (no backport needed). we can change _ABSENT to ABSENT just document it as private type/class in the documentation and not for public use is enough.

So:

current	becomes
NoValue	NO_VALUE
_Absent	ABSENT — leading underscore dropped
_AbsentType	AbsentType
NULL, NO_DEFAULT, NullType, NoValueType, NoDefaultType	unchanged

No deprecation alias. The break is accepted deliberately — NoValue is in pcapkit.corekit.fields.field.__all__ as of #911, so import * callers are affected.

Blast radius, measured (grep -rnw over pcapkit/ and tests/):

NoValue      109 references across 21 files
NoValueType   55 references across 13 files
_Absent       41 references across  5 files
_AbsentType   26 references across  5 files

plus docs/source/pcapkit/corekit/fields/{field,index,misc}.rst, docs/source/contributing/conventions.rst, and the sentinels.rst page that landed in #936.

Three consequences worth handling explicitly, not incidentally:

1. Dropping the underscore makes ABSENT/AbsentType look public to tooling. They must stay out of __all__ — test_sentinel_exports_unit.py::test_the_private_sentinel_is_exported_neither_way pins that, and should keep pinning it under the new names. Privacy becomes documentation-only, per the ruling, so Sphinx will now pick them up where _-prefixed names were hidden; decide whether sentinels.rst documents them as explicitly-private or keeps omitting them. docs(corekit): add the sentinels API page so conventions.rst references resolve (#934) #936 omitted _Absent because it was underscore-private, so that reasoning needs restating either way.
2. conventions.rst needs an object-naming rule added. :152 today states only "Keep the sentinel object's type class naming as <SENTINEL>Type" — it governs the type, never the object, which is the gap that let them diverge. The new rule belongs there.
3. _expected_type_name in test_sentinel_exports_unit.py should collapse. Its bare.isupper() branch and its leading-underscore branch both exist only to cope with the inconsistency; after this, <SENTINEL>Type is one mechanical rule and the helper's special cases become dead.

Blocked on #932, which owns pcapkit/corekit/enum.py (a NO_DEFAULT/sentinel re-export shim), docs/source/contributing/conventions.rst, and tests/protocols/internet/test_mh_unit.py — all three in this change's path. Starts once #932 merges.

[JarryShaw]

Closing: the rename is on main, verified rather than assumed. #939 merged as part of 83c7552b8; main now carries ABSENT = AbsentType() at pcapkit/corekit/sentinels.py:489 and NO_VALUE throughout.

All three renames landed as ruled — NoValue → NO_VALUE, _Absent → ABSENT, _AbsentType → AbsentType, with no deprecation alias, per "take SCREAMING_SNAKE and accept the breaking change (no backport needed)".

The three consequences this issue flagged were each handled rather than left to chance:

• ABSENT/AbsentType are documented as expressly private on sentinels.rst, with the ruling quoted — not omitted. Omission would have been underscore-hiding by another name, which is what dropping the underscore moved away from. Both stay out of every __all__; sentinels.__all__ is ['NULL','NO_VALUE','NO_DEFAULT'], field.__all__ is ['NO_VALUE','Field'], protocol.__all__ is ['ProtocolBase'].
• conventions.rst gained the object-naming rule it previously lacked — it had governed only the type class, which is the gap that let three casings diverge.
• _expected_type_name collapsed to one mechanical rule, with strength preserved by a new regex test pinning ^[A-Z][A-Z0-9_]*$ across all four objects independently of type derivation.

Two findings from the work worth keeping, both reusable: isort reordered five files because NO_VALUE sorts before Field/FieldBase under order_by_type — a pure case change moving imports is not obvious. And pylint's cyclic-import output is nondeterministic: 236 differing lines across two runs of the unmodified baseline, so R0401 churn can never be attributed to a diff.

One thing that was my error, not this work's: I declared tests/protocols/internet/test_mh_unit.py off-limits to the worker after measuring "zero sentinel references in mh.py" — the source, never the test. The test had three, and #939's first head broke all five Python legs with ImportError: cannot import name 'NoValue'. The worker stopped and reported rather than crossing the boundary I drew, which was right.