crypto/tls: replace custom *block with standard buffers
The crypto/tls record layer used a custom buffer implementation with its
own semantics, freelist, and offset management. Replace it all with
per-task bytes.Buffer, bytes.Reader and byte slices, along with a
refactor of all the encrypt and decrypt code.
The main quirk of *block was to do a best-effort read past the record
boundary, so that if a closeNotify was waiting it would be peeked and
surfaced along with the last Read. Address that with atLeastReader and
ReadFrom to avoid a useless copy (instead of a LimitReader or CopyN).
There was also an optimization to split blocks along record boundary
lines without having to copy in and out the data. Replicate that by
aliasing c.input into consumed c.rawInput (after an in-place decrypt
operation). This is safe because c.rawInput is not used until c.input is
drained.
The benchmarks are noisy but look like an improvement across the board,
which is a nice side effect :)
Note that this reduced the buffer for the first read from 1024 to 5+512,
so it triggered the issue described at #24198 when using a synchronous
net.Pipe: the first server flight was not being consumed entirely by the
first read anymore, causing a deadlock as both the client and the server
were trying to send (the client a reply to the ServerHello, the server
the rest of the buffer). Fixed by rebasing on top of CL 142817.