As documented in DirectX7 SDK, d3d7 devices set FPU mode (control word) on every 
call and restore it back to original state before returning, if created in 
DDSCL_FPUPRESERVE cooperative mode. This allows games to work with FPU in a 
possibly incompatible mode and avoid resetting it all the time.

rpcrt4: Fix NdrConformantVaryingStructUnmarshall to use buffer memory if applicable and to reuse memory for embedded pointers.

rpcrt4: Factor out string validation from NDR unmarshalling functions into a separate function, validate_string_data.

Do so by calculating the alignment of members when iterating through
the structures and adding it onto the buffer and memory offsets.

Only call type_memsize once elsewhere in the embedded pointer
processing functions since the return value will be the same from the
second call.

Currently we only check if ARB_HALF_FLOAT_PIXEL is supported. This is
not enough, we need ARB_TEXTURE_FLOAT as well. This fixes some errors
when running the d3d9 visual test with Mesa swrast.