BUG: Avoid precision loss for large integral Decimal values in read_sql with coerce_float=True

[gautamvarmadatla]

Hi all,

This fixes a precision issue in read_sql and read_sql_query when coerce_float=True and a DB driver returns large IDs as decimal.Decimal. Those values can get coerced to float64 and rounded once they’re above 2**53, for example 305184080441754059 becomes 305184080441754048.

The fix is limited to the SQL ingestion path. For object columns, if we see an integral Decimal that would be lossy as float, we convert it to a Python int before dtype inference. Fractional Decimal values still coerce to float as before. Added sqlite-based regression tests for both cases.

• Closes BUG: pd.read_sql is incorrectly reading long int when connecting to Teradata #61667
• Tests added and passed
• All code checks passed
• Type annotations added to new args/methods/functions
• Added an entry to doc/source/whatsnew/vX.X.X.rst (if needed)
• If I used AI, I followed AGENTS.md