Add brotli compressor and decompressor filters

[tobiass-threedy]

Add boost/iostreams/filter/brotli.hpp and src/brotli.cpp, symmetric filters wrapping the Brotli encoder and decoder, modelled on the existing zstd filters. Wired into the CMake and Boost.Build builds, documented in doc/installation.html and doc/bibliography.html, and covered by test/brotli_test.cpp.

The implementation is not mine. It was posted by SSE4 as a work-in-progress patch on the brotli feature request:

#175 (comment)

That patch carried "(C) Copyright Reimar Doeffinger 2024" headers, but the attribution looks like an artefact of how it was produced: after a mechanical zstd->brotli rename, brotli.hpp is 91% line-identical to zstd.hpp and brotli_test.cpp is 83% line-identical to zstd_test.cpp, the three headers match the zstd trio's exact forms with only the year changed from 2018, the "Based on ..." chains were dropped, and even the comment "basically a copy-paste of the gzip test" is inherited verbatim from zstd_test.cpp. So the copyright lines have been restated as SSE4's, over the real derivation chain back through zstd.hpp, lzma.hpp and fusecompress. Only src/brotli.cpp, at 35% shared lines, is substantially new code.

Changes made to the posted patch:

• inflate() reported success whenever it was asked to finish, without consulting BrotliDecoderIsFinished(). Input ending in the middle of a brotli stream therefore decoded as though it were complete. It now raises brotli_error(brotli::truncated_input) when the decoder wants more input and the caller has said there is none, while still letting the caller supply a fresh output buffer when that is all that is missing. Covered by the new truncated_input_test, which fails without this change.

• deflate() only drained the encoder for brotli::finish, so the declared brotli::flush code was never honoured. It now drains for any action other than brotli::run, matching zstd_base::deflate().

• deflate() could also spin forever: it looped while the encoder had more output without checking there was anywhere to put it. It now stops when the output buffer is full so the caller can supply a fresh one.

• brotli_error stored its code as size_t but exposed it as int, which MSVC flags as C4267. Brotli's decoder error codes are a signed enum, so the code is now int throughout, which also removes two casts.

• brotli_base's constructor initialised its members out of declaration order (-Wreorder).

• The CMake integration compiled src/brotli.cpp whenever brotli was found, whether or not BOOST_IOSTREAMS_ENABLE_BROTLI was set, and ran pkg-config unconditionally. It now mirrors boost_iostreams_option(): the option defaults to ON only when brotli is present, forcing it ON without brotli is a hard error, and nothing is compiled when it is OFF. Brotli installs no CMake package configuration file upstream, so detection tries the vcpkg 'unofficial-brotli' package and falls back to pkg-config.

• test/CMakeLists.txt is new to this library, and CMakeLists.txt already had a dormant "if(EXISTS test/CMakeLists.txt) add_subdirectory(test)" guard, so adding the file switched the CMake test path on for everyone while providing only this one test. It now returns early unless both BOOST_IOSTREAMS_ENABLE_BROTLI and Boost::unit_test_framework are present, and registers the test with add_test() so ctest runs it.

Note that "using brotli ;" needs a b2 whose engine provides the "args" module; brotli.jam was added in bfgroup/b2#552 on 2026-05-03 and is not in any released Boost. The CMake build is unaffected.

Tested on Windows against brotli 1.1.0/1.2.0:

• g++ 10.3 and MSVC 19.44 (VS 2022), all six test cases pass under both; truncated_input_test fails when the inflate() fix is reverted.
• CMake: configured and built in all four option states (found/default ON, found/forced OFF, missing/forced ON -> hard error, missing/default OFF).
• Boost.Build: b2 with MSVC compiles src/brotli.cpp into libboost_iostreams and links against the brotli libraries via brotli.jam.