Add primary_key_trigger_spec.rb (mirror of identity_primary_key_spec.rb)

Lock down the `primary_key_trigger:` contract introduced in the
preceding commits. Structure mirrors the existing
`identity_primary_key_spec.rb` (the Phase 1 sibling, #2579) so a
reader who knows one suite finds familiar shape in the other.
Coverage groups:

* Default behavior (no `primary_key_trigger:`) — confirms the
  existing sequence-only path still works and prefetch stays
  enabled. `primary_key_trigger: false` is exercised separately to
  pin the falsy-value contract.

* `primary_key_trigger: true` — exercises the full opt-in:
  sequence + trigger creation, prefetch_primary_key? short-circuit,
  ActiveRecord create! without supplying id (trigger fills via
  NEXTVAL), explicit id supplied (trigger leaves it alone because
  :new.id is non-NULL), and the `:trigger_name` override.

* `add_column` with `primary_key_trigger: true` — pre-2018 routed
  through `create_sequence_and_trigger` so the trigger fired for
  `add_column :foo, :id, :primary_key, primary_key_trigger: true`
  too. This commit's `create_pk_trigger` call site in `add_column`
  preserves that.

* `prefetch_primary_key?` cache invalidation — symmetric with the
  identity suite. Drop-and-recreate flips the verdict for the same
  table name, in both directions.

* Invalid combinations — `id: false`, `id: :integer`, `id: :uuid`,
  composite primary key, and `identity: true` (mutual exclusion).
  All raise `ArgumentError` from
  `validate_primary_key_trigger_options!`.

* Schema dumper roundtrip — the four cells of the matrix
  `{trigger-backed, plain, identity} × {default trigger name,
  custom trigger name}`.

Restored from the 2018 removal (61f8a30), updated for the new API:

* "low-level INSERT with primary_key_trigger: true" — the pre-2018
  "create table with primary key trigger" describe block exercised
  raw `@conn.execute "INSERT ..."` and `@conn.insert(SQL, nil,
  "id")` calls (the connection-level API rather than the AR record
  API). Bringing those examples back catches regressions in the
  RETURNING wiring without going through the full ActiveRecord
  create! path.

* "low-level INSERT — does not raise NoMethodError for :returning_id
  Symbol when logging" — guards the regression reported in #2426
  where the bind-array contained a Symbol that the logging path
  expected to respond to `name`. The fix shipped as part of #2579,
  but the spec was deleted in 61f8a30 alongside the trigger
  feature; restoring it pins the fix against future drift.

* "with non-default :primary_key" / "with non-default
  :sequence_name" — pre-2018 had separate describe blocks for these
  combinations. Brought back in their new shape — both options live
  on `create_table primary_key_trigger: true` (no separate
  `add_primary_key_trigger` call, since this PR intentionally
  doesn't restore that public method).

* "manual trigger creation via raw DDL" — the pre-2018 "create
  triggers" test that verified `@conn.execute` accepts the
  multi-line `CREATE OR REPLACE TRIGGER ... BEGIN ... END;` PL/SQL
  block. Closes the loop on D2 (the design decision to emit the
  trigger as a single PL/SQL block rather than wrapping in
  `EXECUTE IMMEDIATE`): if this test breaks under a future Oracle
  version or driver change, we know the assumption no longer
  holds.

Specs that did not come back from 61f8a30:

* `add_primary_key_trigger` and `has_primary_key_trigger?` —
  pre-2018 public methods that this PR intentionally does not
  restore. The Phase 3 issue (#2597) only specifies the
  `create_table primary_key_trigger:` form; users who wanted the
  separate `add_primary_key_trigger :tbl, options` shape can call
  `connection.send(:create_pk_trigger, :tbl, options)` (with the
  caveat that the private API may change).

* "with separate creation of primary key trigger" — exercised
  `add_primary_key_trigger` inside the same describe; not
  applicable here since we're not restoring that public method.

* schema dumper format `add_primary_key_trigger "test_posts"` —
  the pre-2018 dumper emitted a separate line after `create_table`
  for the trigger. The new dumper emits `, primary_key_trigger:
  true` inline (matching the identity-style pattern), which the
  "schema_dumper roundtrip" describe in this file covers.

Tracked in #2597.

Author: yahonda