Add lbzip2 2.6.5

[caius72]

lbzip2 is a bzip2-compatible compressor. This adds it as lbzip2, levels 1-9, benchmarked single-threaded like the other codecs.

What is vendored

bwt/lbzip2/ holds five sources from lbzip2 2.6.5 — crctab.c, decode.c, divbwt.c, encode.c, parse.c — plus their headers, COPYING (GPL-3.0-or-later) and a README.md recording provenance and how to refresh them. They are byte-for-byte upstream. Upstream's main.c, process.c, compress.c and expand.c are the multi-threaded command-line tool and are not built.

lbzip2_lzbench.c is the lzbench-side wrapper. lbzip2's tool splits work over threads; lzbench benchmarks buffer-to-buffer calls and does its own threading, so the wrapper drives the same low-level codec sequentially, block by block, the way compress.c and expand.c do around lbzip2's scheduler. It also supplies the one function (xmalloc) that decode.c expects from the tool. Registered BENCH_POOL_MT.

Build and portability

• -Ddivbwt=lbzip2_divbwt: zstd's dictBuilder exports a divbwt() too, which collides at link time. -Dxmalloc=lbzip2_xmalloc pre-empts the same problem with a much more common name. Renaming on the compile line keeps the vendored sources unpatched.
• bwt/lbzip2/arpa/inet.h is a shim for MinGW, which has no such header; three of the sources include it for ntohl/htonl alone. It is reachable only through -Ibwt/lbzip2 and defers to the libc macros wherever they already exist.
• No architecture flags are added. The inverse MTF uses a vector shuffle guarded in the source: vqtbl1q_u8 under __aarch64__ && __ARM_NEON, _mm_shuffle_epi8 under __SSSE3__, scalar otherwise. So aarch64 gets the SIMD path automatically, while a default x86-64 build gets the scalar one unless MOREFLAGS=-march=native (or -mssse3) is passed — the numbers below are from an aarch64 machine, where the SIMD path is live.
• make DONT_BUILD_LBZIP2=1 disables it; no CI leg needs excluding.

Correctness

bzip2 -t accepts the output and roundtrip is exact on silesia.tar. Checked with no lzbench errors across all 9 levels, -b1 and -b64 chunking, -T4, incompressible random data, and 1- and 10-byte inputs.

Numbers

Apple M5 Max, macOS 26.5, clang 17, lzbench -ebzip2/lbzip2 -t8,8 -o1c4 silesia.tar. Sorted by compressed size, so each level sits next to its bzip2 counterpart:

Compressor name	Compression	Decompress.	Compr. size	Ratio
lbzip2 2.6.5 -9	43.4 MB/s	103 MB/s	54570631	25.75
bzip2 1.0.8 -9	25.7 MB/s	76.8 MB/s	54572811	25.75
lbzip2 2.6.5 -8	44.0 MB/s	105 MB/s	54660305	25.79
bzip2 1.0.8 -8	25.5 MB/s	77.7 MB/s	54666392	25.79
lbzip2 2.6.5 -7	43.0 MB/s	107 MB/s	55020087	25.96
bzip2 1.0.8 -7	25.9 MB/s	79.0 MB/s	55030188	25.96
lbzip2 2.6.5 -6	43.5 MB/s	110 MB/s	55353765	26.12
bzip2 1.0.8 -6	26.5 MB/s	80.2 MB/s	55382376	26.13
lbzip2 2.6.5 -5	44.5 MB/s	113 MB/s	55682543	26.27
bzip2 1.0.8 -5	26.8 MB/s	81.5 MB/s	55724395	26.29
lbzip2 2.6.5 -4	44.9 MB/s	118 MB/s	56210660	26.52
bzip2 1.0.8 -4	27.3 MB/s	83.3 MB/s	56265716	26.55
lbzip2 2.6.5 -3	44.8 MB/s	125 MB/s	56925346	26.86
bzip2 1.0.8 -3	27.6 MB/s	86.6 MB/s	56992647	26.89
lbzip2 2.6.5 -2	45.0 MB/s	131 MB/s	57997065	27.36
bzip2 1.0.8 -2	27.9 MB/s	88.5 MB/s	58100593	27.41
lbzip2 2.6.5 -1	44.0 MB/s	133 MB/s	60373170	28.48
bzip2 1.0.8 -1	27.7 MB/s	87.6 MB/s	60484813	28.54

Compression is 1.59-1.73x bzip2's and roughly flat across levels, where bzip2 decays from 27.9 to 25.5 MB/s. Decompression is 1.34-1.52x. Compressed size is equal or marginally smaller: 0.004% at -9, 0.19% at -1.

Compression figures are one iteration per level (a single pass over 212 MB already exceeds the 8 s target); decompression got 2-5 iterations, stable to within ~1%.