This upgrade removes some unecessary classpath scanning from
Toucan. This slowness only affected the first query, which would
typically show up as a really slow call to the `/api/user/current`
endpoint (sometimes 10+ seconds).

Previously the spark executor memory was at 1g and Metabase's test
memory was at 3g. This commit lowers Spark's memory to 512m and MB's
memory to 2.5g.

This commit adds query cancelling to the generic sql driver, Druid and
Presto. It cancels the future the query is running in which will
interrupt the query and allow any query cancellation logic to run. If
the database doesn't support query cancellation, or the necessary
information (i.e. query id) isn't available, the code will behave as
it did before.

This commit switches the SQL template parsing to a formally specified
EBNF grammar using instaparse. By specifying it in this way, changes
to the language become more flexible and explicit. This commit also
adds support for multiple parameters specified in an optional clause
as it's baked into the grammar.

Fixes #5492

This commit switches the SQL template parsing to a formally specified
EBNF grammar using instaparse. By specifying it in this way, changes
to the language become more flexible and explicit. This commit also
adds support for multiple parameters specified in an optional clause
as it's baked into the grammar.

Fixes #5492

There are some nice performance improvements in 0.7.4, so updating to
the latest to benefit from them.

The Redshift dependency used previously vendors quite a few
dependencies that we already include in Metabase. By using the
non-embedded Redshift dependency, we can cut about 8 MB in duplicate
JARs and classes.

The embedded Amazon SDK dependency also manipulates the classloader in
a way that triggers warnings on Java 9. This also avoids those warnings.