ci: Restore database backend runs via the generic matrix env field

[krlmlr]

Sync .github/workflows/R-CMD-check.yaml and
.github/workflows/versions-matrix/action.R from cynkratemplate
(cynkra/cynkratemplate#95):
the rcc-full job now applies the generic env field
of test matrix entries to $GITHUB_ENV itself,
before the custom actions run,
because composite actions cannot read the matrix context.

Silent breakage this uncovers and fixes

The backend matrix entries carried a bare test_src field,
and the custom before-install action tried to read it via
${{ matrix.config.test-src }} —
which resolves to an empty string inside a composite action.
The job logs prove it
(run 30139102985, "SQL Server with covr", env of every step: DM_TEST_SRC: empty).

As a result:

• DM_TEST_SRC was never set,
  tests/testthat/helper-src.R fell back to the df source,
  and the six "… with covr" jobs
  tested plain data frames instead of their advertised backends —
  green, but misleading.
• None of the server provisioning in the custom after-install action
  (ankane/setup-postgres / setup-mariadb / setup-mysql / setup-sqlserver,
  keyed on env.DM_TEST_SRC) ever ran.

Fix

• Backend entries now pass DM_TEST_SRC through the generic env field,
  so the after-install action provisions the matching server again
  and the tests really run against SQL Server, Postgres, MariaDB,
  MySQL, DuckDB, and SQLite.
• The entries also set SKIP_UPDATE_SNAPSHOTS=true
  (multi-line env values are now supported by the template),
  keeping the snapshot updater from running the suite a second time
  against the backend and accepting backend-specific snapshots.
  The instrumented-validation entry sets the flag directly, too.
• The custom before-install action loses the env input,
  the apply boilerplate, the broken DM_TEST_SRC step,
  and the DM_VALIDATE-triggered snapshot opt-out —
  all covered by the workflow and the matrix now.

duckdb 1.5.5 storage message

duckdb 1.5.5 announces its extensions-and-secrets location
on driver creation unless the location was chosen explicitly;
on this PR's first CI run the snapshot updater captured that message
(with its session-specific temporary path) into _snaps/dm_sql.md,
and the check step then failed on the path mismatch.
The snapshot is restored to its pre-pollution state here;
the announcement itself is prevented on main by #2487
(~/.duckdb plus DUCKDB_R_HOME in the custom before-install action,
merged into this branch),
so the interim message suppression in the test helpers
has been dropped again.

Heads-up

The six backend jobs have been inert for months;
re-enabling them may surface real drift
(driver installs, backend-specific test failures) in the nightly runs.
That would be true signal, not a regression from this PR.
Note that the nightly run on main currently already has
four unrelated failing jobs (macOS 4.5/4.6, ubuntu-26.04 devel, arm devel).

🤖 Generated with Claude Code

https://claude.ai/code/session_0136C9bRMjBGkUrtLdkYNmAg