Docs and benchmarks for 1.1: single-session Silesia table, graphs, fuzz in CI

[ValisSowilo]

Follow-up to #6: brings the docs and benchmarks in line with 1.1.

Benchmarks

• Silesia preset table re-measured on 1.1 with scripts/bench_session.py (one interleaved session, sentinels at both ends).
  • The main run's opening sentinels read 26% slower than its closing ones, so the first three rows (-7, zpaq, -9) were measured under load.
  • Those three were re-run in a follow-up session bracketed by its own sentinels. They agree to 0.05% (-7) and 0.87% (zpaq), and match the main run's closing readings.
  • The discarded rows are kept in bench_session.json under discarded.
• Every preset is smaller than 1.0.2, by −0.2% to −2.4%. -9 over the whole directory: 35,467,098.
• -1 now beats lpaq1 on every axis: 0.90% smaller, 1.39× faster, 4 MB lighter.
• -5 still beats zpaq -m5 on all three axes: 2.7% smaller, 1.75× faster, 43% less memory.
• -9 is 1.18× zpaq's time here, against 1.07× in 1.0.2's session. A same-session A/B on mozilla (which the transform declines) shows 1.1 is 12% faster than 1.0.2 there, so this is session variance (§26), not the code. The docs say so.
• enwik8 ladder re-measured. 1.1 now beats zpaq -m5 from -3 up; 1.0.2 only did at -7 and -9.
• Graphs regenerated from the new data. bench_session.py gains --exe.

Docs

• README, USAGE and dist/*.txt: tables, marginal exchange rates, memory notes and every derived claim recomputed from the data.
• The "-7 → -9 is no worse than other steps" line is gone: in this session that step costs 19.4× per 1%, and the text now says so, with the cross-session caveat.

Tests

• scripts/wfuzz.py: round-trip fuzz for the word transform. None of the other suites' inputs reach it: they top out at 70 KB, and the transform starts at 256 KB.
• Added to CI on Linux, macOS and Windows, and under UBSan (the UBSan round trip's own input is 238 KB, just under the minimum).

Versions

• Installer script, sign script, CI fallback and dist headers are now 1.1.0.
• The README install guide no longer names a stale installer file.