vkd3d-shader/dxil: Allocate instructions directly in sm6_parser_emit_{alloca,atomicrmw}().
Currently the DXIL parser preallocates instructions in sm6_parser_function_init() and then the helpers just fill the instructions they need and they increment the instruction count by the number of inserted instruction minus one; in the end sm6_parser_function_init() does the last increment if it detects this has to be done. This is uselessly complicated for many reasons, and the goal of this and the following MRs is to fix this by letting each helper just append the instructions it needs. Beside code clarity, the main reason for this is to enable the DXIL parser use the VSIR instruction array iterator abstraction, which will be useful in the future to support forward references, which is AFAIK the last bit we need to fully support SM 6.0.