**Broadcast**
now encodes the wire payload just once and sends the same bytes to
every recipient, before we did one redundant compression per connection.
This primarily reduces CPU load on the server/host.
**Chunking**
Large messages are now compressed before chunking rather than after.
Resulting in slightly smaller payloads.
The receiver now decompresses a single reassembled payload instead of
decompressing every chunk independently, significantly reducing CPU load
on receiving clients.
**Refactor**
Chunking and compression are now low-level wire concerns handled by
Connection rather than being mixed into the high-level message types.
The old `InternalMessageType.Chunk` enum is removed; chunk framing uses
a dedicated wire flag byte alongside `FlagRaw` and `FlagCompressed`.
**Results (Chunking changes)**
Synthetic data, results on real payloads may differ.
Benchmarks (1000 GOs / 2000 components, ~1MB payload, 500 iterations):
Wire size (chunk-first-then-compress): 275KB
Wire size (compress-first): 259KB (5.7% smaller)
Send chunk-first: 0.85 ms/op (old)
Send compress-first: 0.88 ms/op (new)
Recv chunk-first: 1.16 ms/op (old)
Recv compress-first: 0.34 ms/op (new, 3.4x faster)
