Refactoring and improvement of GenerateInt32Stub and related FloatingPointer Helpers

[pengwuibm]

Use this issue to track all code refactoring and improvement for GenerateInt32Stub and related floating point helpers. Observed inefficiency include:

• MOD handling that requires loading double into 2 integer GPRs and then call CallCCodeForDoubleOperation that constructs double out of 2 integer
• The special MOD handling complicates the implementation of LoadNumberAsInt32Double, LoadSmi, LoadNumbers to deal w/ destination of kFPRegisters and kCoreRegisters differently
• Current floating to int conversion uses unoptimized sequence, should use code sequences generated for pwr4 and later architectures
• Unclear why it is necessary to do int32 operation in floating-point (which requires conversion back and forth) on PPC

[pengwuibm]

Simplify the interface for ConvertIntToDouble: 2a7db58

[pengwuibm]

Removed ConvertDoubleToInt and ConvertDoubleToUnsignedInt to use EmitECMATruncate instead 746de9d, which handles corner cases correctly and use more efficient pwr5 instruction sequence for double to int conversion