rpcrt4: The initialisation of [out] variables in the stubless code must be after the unmarshaling of [in] and [in, out] variables has been completed.

This is because the size of an [out] variable could depend on a
conformance value stored in an unmarshalled [in] or [in,out] variable.

rpcrt4: Fix the calculation of element sizes for complex arrays by using a new function ComplexStructSize that doesn't touch the buffer to calculate the size.

Otherwise, this would fail on arrays with a conformance or variance of zero.

shader_get_registers_used is delayed until compile time for some 1.x
shaders, mostly to wait for the right vertex declaration to be
set. This means that on a recompile it will be run again, adding
another instance of each local constant, which in turn causes compile
errors because of constant redeclaration. Just purging the lists
before finding the constants is a simple and reliable solution.

Not all types have a freeing function.

rpcrt4: Add support for calculating the memory size of complex types with embedded non-conformant strings.