* Refactor system tests.

Use parameterized tests for different source formats.

* flake8

* Add additional BQ storage system test fixtures

* Add reader column selection system test

* Add basic reader system test

* Add reader with row filter system test

* Add reading data with snapshot system test

* Add reading column partitioned table system test

* Add system test for column types data conversions

* Add ingestion time partitioned table system test

* Add system test for resuming a read at an offset

* Remove unnecessary protobuf install in noxfile

* Add TODO comment to replace a test helper method

A similar method is planned to be added to the library itself, and when
done, the _add_rows() will not be needed anymore.

* Extract BQ client to session fixture in tests

Creating a client once per system tests session avoids the overhead
of authenticating before each test case.

* Only create BQ storage client once per test run

Creating a client just once avoids the auth overhead on every system
test case.

* Add common credentials fixture for system tests

* Add BigQuery to extra system test dependencies.

* Render with synthtool

* Make all system tests table columns NULL-able

Autodetected schema assumes NULLABLE fields, thus use NULLABLE columns
in test fixture.

* Replace helper function with BQ client method

The _add_rows() helper function is no longer needed, since
client.load_table_from_json() essentially does the same.

* Use explicit schema in data types decoding test

* Adjust a comment related to partial schemas

…ngs (#9333)

Closes #9329.

… (#9738)

…to_arrow + to_pandas, internally (#9997)

Towards https://issuetracker.google.com/140579733

…r (#9994)

* fix(bigquerystorage): resume reader connection on `EOS` internal error

It's infeasible for the backend to change the status of `EOS on DATA`
internal errors, so instead we check the error message to see if it's
an error that is resumable. We don't want to try to resume on *all*
internal errors, so inspecting the message is the best we can do.

* fix: update resumable errors to match string from test failure

* refactor: use more readable any instead of loop

* chore: add split repo templates

* Adjust noxfile

This commit makes sure that bigquery is installed as a stand-alone
package, and not as an editable package from a local path (which does
not exist anymore).

In addition, it makes sure that unit tests are not run with Python 3.8,
as that version is not supported yet.

* reformat synthfile, include regen output

* elide to_grpc_metadata pragma replace

…age (#10)

* feat: update synth to generate v1beta2, v1 endpoints for bigquerystorage

There's also some work to try to accomodate v1alpha2 in this PR, but
there exists further oddities that must be tackled before we can
fully add generation of the alpha client here.

Also, note that this PR does not include manual client modifications
such as streaming offset resumption that are present in the v1beta1
client.  Intent is to address further changes like that in subsequent
PRs.

Co-authored-by: Tim Swast <swast@google.com>

* feat: add manual layer for v1 endpoint

* migrate the manual client/reader work for v1
* TableReferences went away as a first class message, now just a
  formatted string
* changes to read rows
  * estimated_row_count removed due to differences in reported status
  * you no longer have to deal with a StreamPosition message, instead
    there's just a stream name and an offset as top-level request fields
* session creation changes
  * you now provide a prototypical ReadSession message when requesting
    a new read session, and most options (like selected fields, table
    modifiers, and data format) have moved inside it.
  * requested_streams -> max_stream_count

There's additional changes to the surface, but there wasn't much manual
help in front of it so not reflected here.

* updated CHANGELOG.md [ci skip]

* updated setup.py [ci skip]