Fix(athena): Correctly handle PartitionByProperty when it contains Iceberg transforms

[erindru]

This addresses part of SQLMesh issue: TobikoData/sqlmesh#4090

Given the following:

>>> prop = exp.PartitionedByProperty(this=exp.Schema(expressions=[exp.func("bucket", 4, exp.column("foo"))]))

Prior to this PR, when generating against Athena's Trino engine, the value would be truncated:

>>> prop.sql(dialect="athena")
"partitioned_by=ARRAY['bucket']" #should be `partitioned_by=ARRAY['bucket(4, foo)']`

The same value works correctly in Hive mode:

>>> prop.sql(dialect="hive")
'PARTITIONED BY (BUCKET(4, foo))'

However, simply fixing the truncation is not enough. The arguments are swapped when running in Hive mode vs Trino Mode. The argument order is documented in the AWS docs here for Hive and here for Trino.

This PR does the following:

• Fix the truncation in the Presto generator (which is a superclass of Athena's Trino generator). Presto and Trino both support the Iceberg partition transform syntax and the current truncation is a bug for them as well.
• Reorder the arguments for the bucket and truncate Iceberg transforms to be correct for whatever Athena generator is being used (Hive or Trino)

[tobymao]

i don't think this is right, you just want to alter the generator of the property, not do this complex transform

[erindru]

This is just reordering the arguments before it hits the func generator.

I didnt think this logic belonged in the generator for exp.Func because it's very specific:

• It only applies to exp.Func instances that are nested inside an exp.Schema that is itself nested inside exp.PartitionByProperty
• It only applies to Athena
• It only applies to two very specific functions that are used in Iceberg partition transforms, bucket and truncate

[georgesittas]

This looks good to me, although I would clean up a couple of things. If @tobymao's good with getting this in I can do a quick refactor afterwards and share it with you folks.

[erindru]

Closing in favour of a different approach that factors Iceberg partition transforms into their own AST nodes so they can be handled as first-class AST nodes instead of anonymous functions.