Turn them into static functions that take in an hm_fragment. It's not
immediately obvious that the frag_off/frag_len bounds checks and the msg_len
consistency check are critical to avoiding an out-of-bounds write. Better to
have dtls1_hm_fragment_mark also check internally.

Also rework the bitmask logic to be clearer and avoid a table.

Change-Id: Ica54e98f66295efb323e033cb6c67ab21e7d6cbc
Reviewed-on: https://boringssl-review.googlesource.com/3765
Reviewed-by: Adam Langley <agl@google.com>

Notably, drop all special cases around receiving a message in order and
receiving a full message. It makes things more complicated and was the source
of bugs (the MixCompleteMessageWithFragments tests added in this CL did not
pass before). Instead, every message goes through an hm_fragment, and
dtls1_get_message always checks buffered_messages to see if the next is
complete.

The downside is that we pay one more copy of the message data in the common
case. This is only during connection setup, so I think it's worth the
simplicity. (If we want to optimize later, we could either tighten
ssl3_get_message's interface to allow the handshake data being in the
hm_fragment's backing store rather than s->init_buf or swap out s->init_buf
with the hm_fragment's backing store when a mesasge completes.

This CL does not address ssl_read_bytes being an inappropriate API for DTLS.
Future work will revise the handshake/transport boundary to align better with
DTLS's needs. Also other problems that I've left as TODOs.

Change-Id: Ib4570d45634b5181ecf192894d735e8699b1c86b
Reviewed-on: https://boringssl-review.googlesource.com/3764
Reviewed-by: Adam Langley <agl@google.com>

This saves about 6-7k of error data.

Change-Id: Ic28593d4a1f5454f00fb2399d281c351ee57fb14
Reviewed-on: https://boringssl-review.googlesource.com/3385
Reviewed-by: Adam Langley <agl@google.com>

Previously, error strings were kept in arrays for each subdirectory and
err.c would iterate over them all and insert them at init time to a hash
table.

This means that, even if you have a shared library and lots of processes
using that, each process has ~30KB of private memory from building that
hash table.

This this change, all the error strings are built into a sorted list and
are thus static data. This means that processes can share the error
information and it actually saves binary space because of all the
pointer overhead in the old scheme. Also it saves the time taken
building the hash table at startup.

This removes support for externally-supplied error string data.

Change-Id: Ifca04f335c673a048e1a3e76ff2b69c7264635be