fix: connection.Read copies directly from buffer nodes

[xiaost]

The old implementation called Next(n) then copy(p, src). For multi-node reads, Next allocates an intermediate buffer and copies data into it, then copy duplicates it again into p — a double copy plus an unnecessary allocation.

Additionally, Next+Release in Read could invalidate zero-copy slices that the caller still holds from prior Reader.Next calls, since Release frees all consumed nodes from head to read indiscriminately.

Replace Next+copy+Release with readTo, which:

• Copies directly from underlying LinkBuffer nodes into p (single copy)
• Releases fully-consumed nodes inline only when safe (head == node), meaning no outstanding zero-copy references exist
• Skips inline release when head != read, leaving nodes for the caller's eventual Release call

Also fix ioReadWriter to embed io.Reader/io.Writer interfaces and add BUG/deprecation comments for ioReader.Read.

[ppzqh]

假设 head == read，然后通过 Next 单节点读了 read 的一部分。之后 Read 走到这里，好像会把 read 给 release 掉，导致 Next 读的那部分内存回收了。