feat: shim the consumer-side tombstone fix too, until faststream releases it

[aradng]

Stacked on #18 - based on the reverted/working state, not the broken git-dependency one.

ag2ai/faststream#2933 (Optional[Model] = None should resolve correctly for a real tombstone instead of crash-looping on required fields) is open upstream, unmerged, no PyPI release either way. The producer side already gets this treatment via _patch_real_tombstones (0.4.48) - this does the same for the consumer side, so both halves actually work today instead of only the producer.

Three patches, each guarded for idempotency:

• AsyncConfluentParser.parse_message: tags the body with a dedicated TOMBSTONE sentinel (not a byte pattern like b"null"/b"" - can't collide with real message content) when the raw value is a genuine null.
• AsyncConfluentParser.decode_message: maps TOMBSTONE to None.
• faststream._internal.fastapi.route.build_faststream_to_fastapi_parser: a full replacement, since it's a per-subscriber closure factory rather than a patchable class method. Passes a real None through instead of wrapping it as {param_name: None} - that wrapping is exactly what defeats FastAPI's own no-body shortcut (the same shortcut that already correctly resolves Optional[Model] = None for a bodyless HTTP request) and forces field-level validation on every tombstone instead.

Bumps to 0.4.51.

Test plan

• New test in tests/kafka/test_tombstone.py: Optional[Model] = None resolves to None for a real tombstone against a real broker, valid payloads still parse correctly.
• Confirmed the test fails without the shim (assert b'' is None-style failure), passes with it.
• Full tests/kafka/ suite passes (19/19).
• ruff + mypy clean.