Fix PostgreSQL binary-field corruption before encryption

[script3r]

EncryptedBinaryField passed plaintext through the database driver's binary adapter before encryption. With psycopg this encrypted strings such as Binary(b'original\\x00\\xff'), so reads returned corrupted content; memoryview inputs could lose their content entirely in the adapter representation.

Convert buffer inputs directly to bytes and apply the database adapter only to final ciphertext. Add psycopg to the test extra so CI exercises the real PostgreSQL adapter without requiring a server. Runtime dependencies are unchanged.

Validation: 4 real-driver regression cases failed before the fix. All 120 library and 6 example tests pass on Python 3.14 / Django 6.0 / Tink 1.16.1 and Python 3.10 / Django 5.2 / Tink 1.13.0. Covers bytes, bytearray, memoryview, empty buffers, nulls, invalid inputs, and raw SQLite ciphertext. Ruff lint/format and Pyright pass. This is driver-level PostgreSQL testing, not server integration.

Existing correctly stored ciphertext remains readable. Already-corrupted rows need application-specific recovery; the fix cannot reconstruct bytes that were never encrypted.

Stack position: 4 of 6; depends on #9. Land predecessors first and retarget to main as needed.

Landing order: #7 → #8 → #9 → #10 → #11 → #12.

Full review and landing notes. Use merge commits to preserve stack ancestry; squash/rebase merges require rebasing the remaining stack before landing it.