So if "text-compression=br", the .ls file in S3 will get a
Content-Encoding of "br". Brotli appears to compress better than xz
for this kind of file and is natively supported by browsers.

This is necessary for serving log files to browsers.

You can now set the store parameter "text-compression=br" to compress
textual files in the binary cache (i.e. narinfo and logs) using
Brotli. This sets the Content-Encoding header; the extension of
compressed files is unchanged.

You can separately specify the compression of log files using
"log-compression=br". This is useful when you don't want to compress
narinfo files for backward compatibility.

  nix: src/libutil/compression.cc:142: virtual nix::XzSink::~XzSink(): Assertion `finished' failed.

Build logs on cache.nixos.org are compressed using Brotli (since this
allows them to be decompressed automatically by Chrome and Firefox),
so it's handy if "nix log" can decompress them.

We assume that build logs are stored under log/<drv>, e.g.

  /nix/store/q7ab198v13p0f8x8wgnd75dva7d5mip6-friday-devil-0.1.1.1.drv

maps to

  https://cache.nixos.org/log/q7ab198v13p0f8x8wgnd75dva7d5mip6-friday-devil-0.1.1.1.drv

This replaces "nix-store --read-log". It checks the local store and
any configured substituters for the requested logs.

This allows various Store implementations to provide different ways to
get build logs. For example, BinaryCacheStore can get the build logs
from the binary cache.

Also, remove the log-servers option since we can use substituters for
this.

Add missing header <sstream>

Properly set the caFile for aws-sdk-cpp s3

This causes quadratic performance.

This is unnecessary because we make only one connection.

* Unify SSH code in SSHStore and LegacySSHStore.

* Fix a race starting the SSH master. We now wait synchronously for
  the SSH master to finish starting. This prevents the SSH clients
  from starting their own connections.

* Don't use a master if max-connections == 1.

* Add a "max-connections" store parameter.

* Add a "compress" store parameter.

When receiving a very large file, this can prevent the string from
having tobe copied, which temporarily doubles memory consumption.

This allows the NAR to be streamed in the future (though we're not
doing that yet).

When reading a huge string, this halves memory consumption.

(Strictly speaking, this appears only valid in C++17, but who cares...)

Fixes #1173.