Fix SPU output buffer leak

[Ogato]

Abbreviated valgrind output

HEAP SUMMARY:
    in use at exit: ...

8,192 bytes in 1 blocks are definitely lost in loss record 381 of 382
   at 0x8D83828: malloc
   by 0x2E30F0: melonDS::SPU::InitOutput() (SPU.cpp:1085)
   by 0x2E315C: melonDS::SPU::Reset() (SPU.cpp:239)
   by 0x2D98DE: melonDS::NDS::Reset() (NDS.cpp:546)
   by 0x1AEBDE: EmuInstance::updateConsole() (EmuInstance.cpp:1352)
   by 0x1B09C6: EmuInstance::bootToMenu(QString&) (EmuInstance.cpp:1461)
   by 0x1B6403: EmuThread::handleMessages() (EmuThread.cpp:581)
   by 0x1B823F: EmuThread::run() (EmuThread.cpp:445)

LEAK SUMMARY:
   definitely lost: 8,344 bytes in 2 blocks
   indirectly lost: 148 bytes in 2 blocks
     possibly lost: 0 bytes in 0 blocks
   still reachable: 103,561 bytes in 1,134 blocks

The run contains additional Qt/input-method leak noise, but the SPU::InitOutput() allocation is an independent definite leak.

The allocation Valgrind reports is in SPU::InitOutput():

OutputBuffer = (s16*) malloc(2 * newBufferSize * 2);

For the tested 48 kHz output sample rate, newBufferSize becomes 2048, so the allocation size is:

2 * 2048 * 2 = 8192 bytes

That exactly matches Valgrind's 8,192 bytes in 1 blocks are definitely lost report.

SPU::Stop() is not an ownership cleanup path. It only clears and reuses the existing buffer:

memset(OutputBuffer, 0, 2*OutputBufferSize*2);

The owning cleanup path is SPU::~SPU(). Before the fix, the destructor freed AudioLock, BlipLeft, and BlipRight, but did not free OutputBuffer. Therefore the last OutputBuffer allocation made by InitOutput() remained definitely lost when the NDS/SPU object was destroyed.