[Mohith26]

Addresses #369 and the two FIXME Should avoid this dynamic memory call with some sort of pooling comments in CSteamNetworkingMessage::New() / ReleaseFunc().

Every message sent or received does a new/delete of a ~264-byte CSteamNetworkingMessage. This recycles them:

• Each thread keeps a cache of up to 64 free objects, so the common alloc/release is a pop/push with no lock.
• Messages are often allocated on one thread and released on another (received on the service thread, released by the app; or allocated by the app and freed after sending). When a thread's cache is full it moves a batch of 32 to a shared depot, and a thread with an empty cache refills from the depot before going to the heap. The depot is a fixed array (1024 entries) behind a plain std::mutex that is a leaf lock. Anything beyond those bounds is freed as before.
• When a thread exits, its cached objects go back to the depot. GameNetworkingSockets_Kill() frees the depot and the calling thread's cache.
• Off in AddressSanitizer builds (pooling would hide use-after-free) and when the old-Apple-clang thread_local fallback applies; can be disabled with -DSTEAMNETWORKINGSOCKETS_ENABLE_MESSAGE_POOL=0.
• New() also clears m_nUserData now. It was never initialized before (fresh heap memory), but with reuse it would otherwise carry over the previous message's value.

The payload buffer is still a separate malloc, since CSteamNetworkingMessages re-points m_pData/m_pfnFreeData at a header in front of the buffer.

Tests: new message_pool test (added to suite-quick) allocates more messages than the caches hold, scribbles on every user-writable field, and checks recycled messages come back fully initialized, including messages allocated on 4 producer threads and released on 4 consumer threads that then exit. I checked it fails if a field isn't reinitialized. test_connection suite-quick and test_crypto pass on macOS (Apple Silicon, clang, Release), and message_pool/pipe also pass with the pool compiled out.

Benchmark: tests/bench_messages.cpp (built as bench_messages, not run by ctest). Release build, Apple Silicon arm64, 1M iterations, median of 7 runs per process, then median over 13 interleaved processes, pool vs. the same build with -DSTEAMNETWORKINGSOCKETS_ENABLE_MESSAGE_POOL=0:

ns per message	before	after
AllocateMessage(0) + Release()	21.3	4.8
AllocateMessage(64) + Release()	31.6	14.7
alloc 256, then release 256 (64 B)	30.9	16.0
pipe send+recv, one thread, 64 B, batches of 64	78.2	62.5
pipe send+recv, producer and consumer threads	294.5	305.9

The cross-thread pipe case is dominated by lock contention in the pipe path and is within noise (runs ranged 271-314 before, 272-352 after). I haven't measured the UDP path, where allocation is a smaller share of the per-message cost.