[SPARK-13839][SQL] Defer input evaluation and fix Cast issue in IsNotNull filtering for Filter codegen

[viirya]

JIRA: https://issues.apache.org/jira/browse/SPARK-13839

What changes were proposed in this pull request?

This PR improves Filter codegen by:

1. After [SPARK-13751] [SQL] generate better code for Filter #11585, we filter out the rows early if it have null value. So we can only evaluate the inputs used in these filtering before this plan. The evaluation of other inputs can be deferred after the filtering done.
2. Currently we check IsNotNull(a) to find the attributes needed to filter out null values. However, the expression IsNotNull(Cast(a, dt)) will not be checked by this. We should add pattern to find this.

How was this patch tested?

Existing tests.

[SparkQA]

Test build #52997 has finished for PR 11676 at commit b905863.

• This patch fails Spark unit tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #52998 has finished for PR 11676 at commit c539777.

• This patch fails Spark unit tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #53003 has finished for PR 11676 at commit 6e4748b.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[viirya]

cc @davies Can you please review this?

[davies]

We should not add these corner cases here, they should be handled by constraints.

[viirya]

ok. I will send other PR for this.