fix(settings): environment variables beat datajoint.json for every setting

[dimitri-yatsenko]

The module documents environment variables as taking precedence over datajoint.json, but _update_from_flat_dict enforced that only for keys in the hand-kept ENV_VAR_MAPPING. For every other setting the file replaced the value read from the environment: jobs.* (DJ_JOBS_*), provenance.* (DJ_PROVENANCE_*), database.use_tls, database.reconnect, and the top-level safemode, query_cache, download_path and ignore_config_file. A deployment could not, for example, turn off safemode or provenance capture for one worker when the project's file set it.

ENV_VAR_MAPPING is now derived from the settings models: a field's variable is its validation_alias, otherwise its model's env_prefix plus the field name. A setting added later is covered as soon as it is declared, including one that has only a validation_alias. Models without an env_prefix (connection, display) contribute only their aliased fields, matching the documented variables, and stores stays with _apply_stores_env, which already runs after the file. The eleven names in the old mapping are unchanged.

The tests list every setting that reads an environment variable, assert the derived mapping matches that list, and check for each one, in both flat and nested file layouts, that pydantic reads the variable and the file does not replace it.

Closes #1585.