I've run across problems like this when writing parsers. The way I would do it is to liberalize the grammer and leave the checking of the ident to a later processing stage. For instance, I would probably use something like:

ident : IDENT | INDENT_QUOTED | ident '.' IDENT | ident '.' IDENT_QUOT
+ED ;
[download]

and not have the grammar distinguish between table_ident, column_ident and database_ident. The checking of the proper kind of ident can be done later. This approach also generally makes it easier to recover from errors.

Update: The reduce-reduce conflict is this: when trying to parse a column_ident and the current input token is ident with look-ahead of ., the following reductions are possible:

  database_ident -> ident # i.e. column_ident -> ident '.' ident '.' i
+dent
  table_ident -> ident    # i.e. column_ident -> ident '.' ident
[download]

One solution would be to reverse the recursion:

database_ident: ident ;
table_ident: ident | ident '.' database_ident ;
column_ident: ident | ident '.' table_indent ;
[download]

Then when the look-ahead is '.' the grammar knows exactly what production to try to complete.