The source of the dependency bug is in old versions of google-cloud-core that
depend on too-old versions of google-api-core.

Some tests are still failing, but
we're far enough along that we have the right shape, I think.

The moral equivalent of    "where foo in (@bar)", where bar is an array
which actually need to be  "where foo in unnest (@bar)".

Doh!

Other fixes:
- Handle BIGINT
- Fix string leteral formatting (and start type-specific adaptations).

…ures

The SQLAlchemy like convenience functions (e.g. ) escape incorrectly for
BigQuery, so re-escape.

We could make that work, if we want to. :)

…nt_digits

…ry keys, etc., that BigQuery doesn't have.

#78

…xpect

The source of the dependency bug is in old versions of google-cloud-core that
depend on too-old versions of google-api-core.

So we don't have t mock at the api level.

#78

- Run tests in temporary directory rather than sharing memory connections.
  Because simpler. :)
- Introduce cross-connection state and record queries in it, so tests can
  make assertions bout generated queries.

…laceholder

The BigQuery Python Client supports an extended placeholder syntax
that includes type information, as in `%(foo:INT64)s` (named) or
`%(:INT64)s` (unnamed).

When we know the type, include it in the placeholder.

The numeric tests now tun since we started passing type info from sqla to bigquery.

The CTE tests test features that don't exist in bigquery.

Although the test isn't actually testing dialect code.  Maybe it
should be skipped instead.

Also set the profile test pasth to a more reasonable value, although
it doesn't seem to be used. <shrug>

Also inlined colspecs code, because there wasn't much and it
facilitated separating literal processing into a function.

It didn't buy us anything.

Because:

- It duplicated the `--dropfirst` option.

- Unlike `--dropfirst`, it didn't honor the `--dburi` option that lets you use an alternate URL/dataset.

The schema-tests schema is hard coded, and cleaning it and running it's tests could conflict with other test runs.

Also, use multiple simple --only-rerun options rather than one complex one.