Restore shared HTTP grants after extension lifecycle tests

[RKS (rksharma-owg)]

12_extension_lifecycle.sql can pass while leaving df_e2e_user unable to use HTTP. The shared setup grants HTTP access, but the lifecycle test drops/recreates the extension and restores privileges with the default include_http => false. A later HTTP test then depends on whether setup has run again.

Restore the same HTTP access as 00_setup_playground.sql in the final re-grant. Add end-of-test assertions for both df.http() and df.http_multipart() privileges. The existing checks that administrative helpers remain inaccessible still run.

Related to #381. This addresses one reproduced instance of shared state leakage; it does not provide general per-test database isolation.

Validation

Tested in disposable Docker containers on Linux arm64, using source-built pg_durable v0.2.8 with http-allow-test-domains:

Check	PostgreSQL 17.11	PostgreSQL 18.6
New regression before the grant fix	Fails at the HTTP privilege assertion	Fails at the HTTP privilege assertion
00_setup_playground.sql and 01_core_primitives.sql	Pass	Pass
Updated 12_extension_lifecycle.sql, twice in the same database	Pass both times	Pass both times
Non-superuser df.http() constructor after each lifecycle run, without repairing grants	Pass both times	Pass both times

SQL files were run with docker exec "$container" psql -X -U postgres -v ON_ERROR_STOP=1 -f /repo/tests/e2e/sql/<file>. The constructor probe used SET SESSION AUTHORIZATION df_e2e_user; SELECT df.http('https://httpbingo.org/get', 'GET'); in a separate connection; it makes no network request.

cargo fmt -p pg_durable -- --check and git diff --check passed. This changes only the E2E SQL file; the full suite was not run. The local Docker builds used two Cargo jobs, thin LTO, and eight release codegen units.

AI assistance: implemented and validated with Codex.