1) Use `assert` and `refute` where possible.
2) Separately include the setup, subject, and assertions in each test - don't hide the tested call in an assertion method.
3) Name things based on their role rather than incidental facts about them - e.g. `@bird[1]` -> `bird_to_destroy`. `bird2_deletion_attributes` -> `destroy_bird_attributes`.
4) Use more succinct naming where possible - e.g. `birds_with_callback` -> `birds_with_add`, `@pirate_with_two_birds` -> `@pirate`

Restore the use of `#add_to_target` for nested attribute updates on existing records, and don't bother updating the association if the update is going to be rejected anyway.

This requires adding a `skip_callbacks` argument to `#add_to_target`
so that we don't call the callbacks multiple times in this case,
which is functionally an application of existing association data,
rather than an addition of a new record to the association.

Issue #1: :before_add callback is called when nested attributes assignment assigns to existing record if the association is not yet loaded
Issue #2: Nested Attributes assignment does not affect the record in the association target when callback triggers loading of the association

If you're using an in-memory database for the test database,
reconnecting will lose all information that schema:load did for us.

If 'enable_extension' is used in a migration's 'change' method, use
'disable_extension' on down migration (and vice-versa).

This commit fixes the build but we need to further investigate the root cause,
to see if it can be fixed.

The test fails under `MysqlAdapter` with:

```
ActiveRecord::UnknownAttributeError: unknown attribute: なまえ
```

The root cause is that the returned column names from the adapter
have the wrong encoding:

```ruby
columns = Weird.columns_hash.keys
columns # => ["id", "a$b", "\xE3\x81\xAA\xE3\x81\xBE\xE3\x81\x88", "from"]
columns.map(&:encoding) # => [#<Encoding:ASCII-8BIT>, #<Encoding:ASCII-8BIT>, #<Encoding:ASCII-8BIT>, #<Encoding:ASCII-8BIT>]
columns.map {|c| c.dup.force_encoding "utf-8"} # => ["id", "a$b", "なまえ", "from"]
```

From what I can tell the MySQL variables are fine:

```
ActiveRecord::Base.connection.execute("show variables")

...
character_set_client utf8
character_set_connection utf8
character_set_database utf8
character_set_filesystem binary
character_set_results utf8
character_set_server utf8
character_set_system utf8
character_sets_dir /usr/local/Cellar/mysql/5.5.29/share/mysql/charsets/
collation_connection utf8_general_ci
collation_database utf8_unicode_ci
collation_server utf8_general_ci
...
```

Closes #11062

Fix in-memory tests

[ci skip]

This change was breaking the Railties tests. See
https://travis-ci.org/rails/rails/jobs/9865969

The deprecation message was removed on 50cbc03d but the code was not.

If the `db/` directory is not present on a remote machine it will blow up in unexpected ways with error messages that do not indicate there is a missing directory:

```
SQLite3::CantOpenException: unable to open database file
```

This PR checks to see if a directory exists for the sqlite3 file and if not creates it for you.

This PR is an alternative to #11692 as suggested by @josevalim

`result_metadata` returns a new object each time it is called, so calling `result_metadata.free` is essentially a noop.  Instead call `free` directly on the metadata when we're done with it.

When mixing postgresql and another adapter like sqlite3 (for dev and test
respectively), the task `db:test:prepare` will fail due to the `enable_extension`
method not being defined in the abstract adapter. This patch simply adds an
empty definition to prevent it.

Closes: #11706

tesetcases assertion to case insensitive because Oracle database adapter
handles table name in uppercase.