[SQL Server] Inline DEFAULT_REPLACEMENT_STRING to reduce impact on 2100 query parameter limit #1840

Problem

I had a custom running random row validations on SQL Server. Random row values are bound into a query using query parameters. If there are multiple columns in a primary key then the 2100 query parameter limit can quickly be reached. I suggested the customer use the formula:

random row batch size = (2100 / num pk columns)

However this was not adequate because, in hash validation, DVT constructs a coalesce(cast__<col>, 'DEFAULT_REPLACEMENT_STRING') expression for every nullable column and SQLAlchemy compiles every column's 'DEFAULT_REPLACEMENT_STRING' as a separate bound ? parameter. Therefore, if a table had 1000 nullable columns the maximum random row batch size was cut in half.

Proposed Solution

Inline DEFAULT_REPLACEMENT_STRING as a SQL literal:
Render the constant replacement string inline in SQL (e.g., via sa.literal_column("'DEFAULT_REPLACEMENT_STRING'") or literal_binds in the fillna compilation path) rather than binding a ? parameter per column. This removes hashed_column_count from the parameter count across all SQLAlchemy-backed engines.

We should also validate the DEFAULT_REPLACEMENT_STRING token (which I believe can be overridden in a YAML config file) to protect against SQL injection.