This provides correct rounding in the IEEE 754 sense for the Value attribute
of floating-point types in more cases, by bumping the number of significant
bits used in the initial integer mantissa obtained from parsing.

Tested on x86_64-pc-linux-gnu, committed on trunk

gcc/ada/

	* libgnat/s-valuer.ads (System.Value_R): Add Parts formal parameter
	as well as Data_Index, Scale_Array and Value_Array types.
	(Scan_Raw_Real): Change type of Scale and return type.
	(Value_Raw_Real): Likewise.
	* libgnat/s-valuer.adb (Round_Extra): Reorder parameters and adjust
	recursive call.
	(Scan_Decimal_Digits): Reorder parameters, add N parameter and deal
	with multi-part scale and value.
	(Scan_Integral_Digits): Likewise.
	(Scan_Raw_Real): Change type of Scale and return type and deal with
	multi-part scale and value.
	(Value_Raw_Real): Change type of Scale and return type and tidy up.
	* libgnat/s-valued.adb (Impl): Pass 1 as Parts actual parameter.
	(Scan_Decimal): Adjust to type changes.
	(Value_Decimal): Likewise.
	* libgnat/s-valuef.adb (Impl): Pass 1 as Parts actual parameter.
	(Scan_Fixed): Adjust to type changes.
	(Value_Fixed): Likewise.
	* libgnat/s-valrea.adb (Need_Extra): Delete.
	(Precision_Limit): Always use the precision of the mantissa.
	(Impl): Pass 2 as Parts actual parameter.
	(Exact_Log2): New expression function.
	(Integer_to_Real): Change type of Scale and Val and deal with a
	2-part integer mantissa.
	(Scan_Real): Adjust to type changes.
	(Value_Real): Likewise.