test: deflake sibling wall-clock perf-budget tests (pubmed / sqlcipher / alembic)

PR #6623 fixed the wall-clock contention flake for the three scrub_error budget tests (test_log_sanitizer.py::TestScrubErrorUrlQueryRedaction) by measuring CPU time (time.process_time) — after #6612 and two later CI runs failed test_scrub_error_is_bounded_on_a_run_of_bare_schemes on wall-clock noise alone. The same fragility exists in sibling budget tests elsewhere in the suite, which will keep flaking on loaded xdist runners:

Test	Budget	Note
tests/research_library/downloaders/test_pubmed.py:332	~0.1s wall	near-zero headroom; one GC pause kills it — worst offender
tests/security/test_sqlcipher_integration.py:1722	1.0s perf_counter over real SQLCipher I/O	I/O-bound: cannot switch to process_time — needs a widened wall budget instead (see #6623's body for the factor analysis: core variance ×1.3, pollution ×1.7)
tests/db/test_alembic_migrations.py (4 asserts)	5–30s wall over real migration I/O	modest headroom on loaded CI

Suggested treatment per class:

• CPU-bound budgets → switch to time.process_time() and size the budget from measured healthy cost × worst-case compounding (see the convention note in TestScrubErrorUrlQueryRedaction's docstring).
• I/O-bound budgets → keep the wall clock but widen with measured headroom; consider whether the budget guards a property that a size-invariant ratio (t(2n) < 3·t(n)) would guard better.

Found during the 5-round review of #6623; a repo-wide sweep is git grep -nE 'assert.*(elapsed|took|duration).*<' -- tests/.