Fixes #224

This reverts commit 48ba0b8d.

This commit adds special handling for the very first Initial packet
which is corrupted and discarded.  If server receives such packet, it
has to drop connection silently without sending any CONNECTION_CLOSE.
Then valid packet might come later.  Keeping connection state is
harmful because corrupted packet has wrong SCID, which drops later
valid packet because mismatched SCID.  New error code
NGTCP2_ERR_DROP_CONN is added which tells server application to drop
connection silently.

Use longer timeout because packet loss rate is incredibly high (30%).

Clarify the behaviour of ngtcp2_conN-writev_stream with
NGTCP2_WRITE_STREAM_FLAG_MORE.

Remove handling of NETWORK_ERR_SEND_BLOCKED in example server

Fix typo

The example server uses sendmsg() in blocking mode and NETWORK_ERR_SEND_BLOCKED is never returned as an error.

Remove conn and user_data from encrypt/decrypt/hp_mask callbacks so
that they can be used without conn.

Bring back the processing of Handshake packets.  Doing so can reset
idle timer and reset PTO count which is useful to deal with packet
loss during handshake.

With this change, ngtcp2 library does not increase remote stream limit
automatically.  Application should extend the stream limit by calling
ngtcp2_conn_extend_max_streams_bidi or
ngtcp2_conn_extend_max_streams_uni.  Automatic increment is
problematic because the application stream might be closed well after
QUIC stream is closed (e.g., lost packets for QPACK stream might delay
HTTP level stream closure).

The rule is that if ngtcp2_stream_open is called, then application is
responsible to extend the stream limit.

Remove redundant Buffer::begin

Stream must be closed in nghttp3 callback because quic stream close
comes before http stream close because of packet reordering or loss.

It always equals to buf.data().