[snarfable]

Replace the linear, memmove-compacting connection buffer with a
fixed-capacity single-producer/single-consumer ring buffer.

Buffer

• Power-of-two capacity; indices are monotonically increasing counters
  masked on access, so full and empty states are unambiguous.
• Lock-free: the producer publishes with a release store to head, the
  consumer frees space with a release store to tail; the opposite index
  is read with acquire. Each side caches the other's index and refreshes
  it only when the buffer looks full/empty, and head and tail live on
  separate cache lines to avoid false sharing.
• New API: Write, Space, Peek, Read, ContiguousData (zero-copy view up
  to the wrap point), Consume, Drain, Available, Capacity, Empty, Full.
  Shift() is kept as an alias for Consume().
• Non-copyable/non-movable (owns atomics).

Fixes in the old Buffer

• Write() copied every chunk to offset 0, overwriting any unparsed bytes
  left from a partial AMQP frame.
• Write() refused a write that would exactly fill the buffer.

IConnectionHandler

• Feed AMQP::Connection::parse() directly from ContiguousData(); when
  the readable bytes wrap the ring, linearize them into a scratch buffer
  so frames are always contiguous.
• Only read from the socket as many bytes as the ring can hold.
• Check receiveBytes() for errors instead of passing a negative count to
  Write(); fix savail being unsigned so the < 0 error check can fire.
• _sendDataFromBuffer() now consumes the bytes that were actually sent.

Tests and benchmark

• tests/buffer_t.cpp: capacity rounding, append semantics, full/empty,
  peek/read/consume/drain, wrap-around, contiguous views, a randomized
  20,000-operation comparison against a reference model, and two
  producer/consumer thread stress tests (copying and zero-copy paths).
  Clean under ThreadSanitizer and AddressSanitizer/UBSan.
• benchmarks/buffer_bench.cpp: SPSC throughput across chunk sizes,
  enabled with -DBUILD_BENCHMARKS=ON.

[snarfable]

Good implementation of a ring buffer.