rsx fp/vp analyzers: Fix strict type aliasing and improve codegen

[elad335]

• Obey strict type aliasing rule in rsx fragment/vertex program analyzers.
• Optimize instructions reads using SSE loads instead of using references/pointers to specific instruction members at a time. (memcpy 16 bytes is compiled to unaligned 128-bit SSE2 loads and stores)
• Remove lots of reinterpret casts, make the code easier to read.
• Replace qword type with the much more useful v128 type which provides extra utilities such as builtin operators ==, != for instruction comparisons.

[kd-11]

There are some lingering concerns, but I'm already working on shaders anyway and can fix anything that crops up, there's no need to hold up the PR itself.