WL#606: Estimate rows for deactivate_non_unique_index()

LOAD DATA INFILE, CREATE TABLE ... SELECT and INSERT ... SELECT should
estimate the number of rows to be inserted in the table and call
start_bulk_insert() with this number.

For LOAD DATA INFILE one can do this with:

file_length / (min-length-of-packed-row)

For CREATE TABLE ...SELECT and INSERT ... SELECT this means that this should
be done in ::initialize_tables() instead of in ::prepare() as we in the
later case have an idea of how many rows there could be in the result set.

(Can be stored in join->number_of_records, in sql_select.cc::find_best(),
when we update join->best_read)

Also, start_bulk_insert() now disables indexes only if there are no
rows in the table, otherwise bulk_insert_tree is used. In fact, indexes could be
disabled even when table is not empty, if it will give a speedup.