Make dbinit.sql slightly less order-dependent

[smklein]

This PR does two things:

1. It avoids checking for the order-dependent NOT NULL constraint in information schema, opting to parse the output of SHOW CONSTRAINTS and information_schema.columna.is_nullable instead.
2. It re-orders the db_metadata creation to the end of dbinit.sql (mostly for clarity)

Fixes #4286

[gjcolombo]

It re-orders the db_metadata creation to the end of dbinit.sql (mostly for clarity)

Hooray!

I've only skimmed the PR but will ask anyway: If this is now legal, is the guideline in the schema README to add new tables only to the end of dbinit.sql still relevant? (If it's not, we should remove it and consider replacing it with guidance that db_metadata should always be last.)

[smklein]

It re-orders the db_metadata creation to the end of dbinit.sql (mostly for clarity)

Hooray!

I've only skimmed the PR but will ask anyway: If this is now legal, is the guideline in the schema README to add new tables only to the end of dbinit.sql still relevant? (If it's not, we should remove it and consider replacing it with guidance that db_metadata should always be last.)

Sure thing, I just pushed a small update to the schema/crdb README.

[david-crespo]

Looks good. I confirmed this change makes the schema tests pass in #4261.

[bnaecker]

I don't have enough context to really approve, but I don't see any issues!