Refuse separator-bearing env secret ids only in multi-team mode

[potiuk]

Another follow-up after @amoghrajesh's #71078, which fixed this same regression in the AWS, Key Vault, Google Secret Manager and Lockbox secrets backends. The core EnvironmentVariablesBackend carries it too, and was not covered there — cc @vatsrahul1001 @amoghrajesh.

The backend refuses any connection or variable id carrying the team namespace separator, because such an id cannot be attributed to a team unambiguously — a team name may itself contain the separator, so _a___b___c reads as team a with id b___c just as well as team a___b with id c.

That refusal fires unconditionally, including in deployments that never enabled multi-team mode. There, an ordinary connection or variable id that happens to contain ___ is silently reported as missing — no exception, and nothing distinguishes it from a secret that was never set.

No team scoped secret can be read while multi-team is off: the lookups are only ever handed a team_name in that mode. So the id has a single meaning there and nothing needs refusing. Same gate as #71078, applied at the same place in the logic.

What this changes for the tests

The existing team-scoping tests all assert the refusal, so they now run with multi-team enabled. Two of them (test_connection_env_var_do_not_access_team_specific and test_variable_env_var_do_not_access_team_specific) would otherwise fail, because AIRFLOW_CONN__TEAM___TEST_MYSQL legitimately resolves as an ordinary connection named _team___test_mysql once multi-team is off — that is the intended trade, and the same one #71078 makes on the provider side.

Test plan

• New TestEnvironmentVariablesBackendWithoutMultiTeam — verified against unmodified code: it fails with assert None == 'team-agnostic-value', the reported symptom exactly
• test_secrets.py, test_secrets_environment_variables.py, test_secrets_backends.py — 72 passed
• Full airflow-core/tests/unit/always — 2008 passed; the one failure (example_mysql_to_gcs.py) is a missing mysqlclient on the host and reproduces on unmodified main
• ruff / ruff format / static checks clean

No newsfragment: #70736 is milestoned 3.3.1 and unreleased, so the regression never reached users.

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 5)

Generated-by: Claude Code (Opus 5) following the guidelines

[potiuk]

Closing in favour of #71079