Fix!: bump sqlglot to v25.29.0, fix info schema view handling in bigquery

[georgesittas]

Fixes #3317

The goal of this PR is to enable SQLMesh to correctly handle information schema view references in BigQuery. The main problem with those until now was that, in their fully-qualified form, they comprised 4 identifiers:

project.dataset_or_region.INFORMATION_SCHEMA.SOME_VIEW

This means that we'd end up with Table references of mixed nesting, e.g. model names comprise 3 identifiers:

project.dataset.model_name

Mixing multiple nesting levels in table references is prohibited by SQLGlot's schema module [1, 2], in order to avoid issues related to ambiguity. So, one workaround for that was to represent information schema views using 3 identifiers at parse time, only for BigQuery. Other engines don't allow >3 identifiers in their table references based on my investigation.

I went with this approach because making the schema module more lenient, i.e. allowing multiple nesting depths, was quite complex. We rely on the invariant that the depth is the same in several places and the scope of the current approach seemed way smaller in comparison.

I guess one thing we'll need to be careful about is that parsing BigQuery's information schema views without specifying the dialect can result in an incorrect AST representation, because we represent the first example with the 4 identifiers using a Dot instead of merging the last two parts into a single Identifier, as done in BigQuery's parser.

For additional context, please refer to:

• Fix: enable fetching schema for models querying INFORMATION_SCHEMA
• Fix(bigquery)!: parse information schema views into a single identifier

[erindru]

I'm.. amazed this works