Always generate MemoryAccessAlignedMask for loads/stores

[olvaffe]

half4 loads/stores are aligned to 8 bytes. Without MemoryAccessAlignedMask to explicitly specify the alignment, vulkan drivers might need to assume the worst case which is sizeof(half) or VkPhysicalDeviceLimits::minStorageBufferOffsetAlignment.

In the case with radv (mesa amd driver), instead of generating 2 buffer_load_dwordx4 to load a 4x4 half4 matrix, it would generate 4x buffer_load_dwordx3 plus bit manipulations instead.