fix(optimizer)!: Preserve DPipe in simplify_concat

[VaggelisD]

Fixes TobikoData/sqlmesh#2439

The simplify_concat rule would merge a list of binary exp.DPipe concatenations into one exp.Concat function call while constant folding any running string literals on the way. That (optimized) expression would then be generated as CONCAT(a, b, ...) which might not have the same semantics as a || b in every dialect e.g. tsvector in Postgres.

This PR aims to preserve exp.DPipe by tagging any exp.Concat expressions generated by simplify_concat with dpipe_source=True, which will cause them to be generated back as a list of (now optimized) binary exp.DPipe concatenations

[georgesittas]

Isn't the issue local to simplify_concat? Why do we need to introduce yet another arg to Concat for this? Imo this should be as simple as checking the type of the input expression and then ensuring we return that type before returning, eg. if we have a Concat at the end but the original type was DPipe, then do something like this (without the sql part):

sqlglot/sqlglot/dialects/dialect.py

Lines 875 to 876 in 4f1691a

	def concat_to_dpipe_sql(self: Generator, expression: exp.Concat) -> str:
	return self.sql(reduce(lambda x, y: exp.DPipe(this=x, expression=y), expression.expressions))

[VaggelisD]

@georgesittas Check out the latest commit, reduce did simplify things a lot, thanks!

[georgesittas]

LGTM