* Don't erase __attribute__ under clang-cl

The shim fires when __GNUC__ is absent, which clang-cl does not define even
though it supports __attribute__ fully. Erasing it breaks clang's own intrinsics
headers: mmintrin.h declares __m64 as a vector via __attribute__((vector_size)),
so it degrades to a scalar and every use of it fails to compile.

* Don't erase __attribute__ under clang-cl

clang-cl supports it but does not define __GNUC__, so the shim fired and broke
clang's own intrinsics headers.

Expand the main `GfxPool` buffers to 1MiB

The decoder cached on a note was keyed on nothing, so a note reused for a
different streamed sample carried on decoding the previous track. Audible as
the wrong custom music: the sequence and soundfont the audio editor reports
are correct, only the samples reaching the mixer are not.

Reproduced on the game-start cutscene chain with a streamed music pack in 4
of 6 runs; 0 of 6 after.

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: serprex <159546+serprex@users.noreply.github.com>

Avoids timer from Volvagia carrying over to DMC

Root cause of three identical field crashes (audio thread, opcode fetch
through a pointer with its low 32 bits overwritten, seconds after scene
transitions): AudioHeap_AllocPermanent writes permanentCache[index] with
index = permanentPool.count and no bound against the 32-entry array. In
SoH every soundfont sync-load is forced permanent, and custom sequences
whose SEQ.xml says CachePolicy="Temporary" ALSO allocate permanently
(the factory stores the LUS enum where CACHE_TEMPORARY == 0, while
AudioLoad_SyncLoad's switch reads 0 with the ROM convention
'permanent'). A pack with ~60 streamed customs plus vanilla fonts pushes
count past 32 within a session, after which each allocation sprays a
{ptr, size, tableType/id} triplet at 24-byte stride through
gAudioContext - entry[135]'s ptr field lands exactly on
seqPlayers[0].scriptState.pc and entry[156] on seqPlayers[1]'s (both
verified against the crash-dump registers).

- permanentCache raised 32 -> 512 (12 KB) and AllocPermanent refuses
  allocations past the array instead of corrupting memory.
- Same unbounded-index disease fixed in the three sibling writers:
  AllocCached's persistent path (16-entry array; CACHE_EITHER degrades
  to temporary, hard persistent requests fail cleanly),
  AllocPersistentSampleCacheEntry, AllocTemporarySampleCacheEntry.
- seqLoadStatus malloc sized for the full id space (sequenceMapSize +
  0xF) matching sequenceMap; custom ids above sequenceMapSize previously
  overflowed the allocation by up to 15 bytes.

Upstream SoH bugs, not branch-introduced - this branch's many-track
packs merely made the overflow reachable in normal play. Standalone
upstreamable fix.

Co-authored-by: serprex <159546+serprex@users.noreply.github.com>