minecraft: cut per-batch decode allocations ~1500x with pooled decompression

[HashimTheArab]

Reading a compressed batch used to allocate the full decompressed payload plus a packet collection on every read. The read loops now decompress into a pooled buffer and allocate ~50 bytes per batch instead of ~74 KB, and a single queued packet no longer pins the entire decompressed batch's backing array.

Benchmark (64 packets × 1 KiB in one flate batch, Apple M3 Pro, three runs):

Read path	Time	Allocated
flateCompression.Decompress (old)	19.3–21.0 µs/op	~74 KB/op, 2 allocs/op
Decoder.DecodeFunc (new)	16.7–19.7 µs/op	48–50 B/op, 1 alloc/op

Changes

• Add Decoder.DecodeFunc, which calls a callback for each packet in a batch with slices borrowed from a pooled decompression buffer. The batch is validated in full before the first callback, so a malformed batch dispatches no packets, preserving the all-or-nothing behavior of Decode.
• Switch the dialer and listener read loops to DecodeFunc. Packets are copied only at the two points where they can outlive the batch (the Conn.packets channel and deferPacket), via packetData.ensureOwned.
• Add append-style decompression for flate and snappy behind an unexported appendDecompression interface, so batches decompress into an existing pooled buffer. Buffers above 1 MiB are dropped instead of returned to the pool.
• Decoder.Decode keeps its API and behavior. Its returned packets are now backed by one caller-owned allocation, where they previously aliased the decoder's internal read buffer and were only valid until the next Decode call.
• PacketFunc now receives a copy of the payload, so it can retain it safely.

Compression-limit handling is untouched; this is orthogonal to #503.

Validation

• go build ./..., go vet ./..., go test -race -count=1 ./...

[HashimTheArab]

ive been running this for a couple of days now and its running great, no issues