Misc: Align CPU state and SPU voices to cache lines

Fixes the ~12% performance regression from the texture replacement
namespace'ify commit. Apparently LTO was placing the CPU struct in the
middle of a cache line...

src/common/types.h

@@ -199,6 +199,7 @@ static constexpr u32 HOST_CACHE_LINE_SIZE = 128; // Apple Silicon uses 128b cach
 #else
 static constexpr u32 HOST_CACHE_LINE_SIZE = 64; // Everything else is 64b.
 #endif
+#define ALIGN_TO_CACHE_LINE alignas(HOST_CACHE_LINE_SIZE)
 
 // Enum class bitwise operators
 #define IMPLEMENT_ENUM_CLASS_BITWISE_OPERATORS(type_)                                                                  \

src/core/cpu_core.cpp

@@ -89,7 +89,7 @@ static bool WriteMemoryByte(VirtualMemoryAddress addr, u32 value);
 static bool WriteMemoryHalfWord(VirtualMemoryAddress addr, u32 value);
 static bool WriteMemoryWord(VirtualMemoryAddress addr, u32 value);
 
-State g_state;
+alignas(HOST_CACHE_LINE_SIZE) State g_state;
 bool TRACE_EXECUTION = false;
 
 static fastjmp_buf s_jmp_buf;

src/core/cpu_core.h

@@ -121,7 +121,7 @@ struct State
   static constexpr u32 GTERegisterOffset(u32 index) { return OFFSETOF(State, gte_regs.r32) + (sizeof(u32) * index); }
 };
 
-extern State g_state;
+ALIGN_TO_CACHE_LINE extern State g_state;
 
 void Initialize();
 void Shutdown();

src/core/spu.cpp

@@ -401,7 +401,7 @@ static std::array<std::array<s16, 128>, 2> s_reverb_downsample_buffer;
 static std::array<std::array<s16, 64>, 2> s_reverb_upsample_buffer;
 static s32 s_reverb_resample_buffer_position = 0;
 
-static std::array<Voice, NUM_VOICES> s_voices{};
+ALIGN_TO_CACHE_LINE static std::array<Voice, NUM_VOICES> s_voices{};
 
 static InlineFIFOQueue<u16, FIFO_SIZE_IN_HALFWORDS> s_transfer_fifo;