docs: rewrite the README for the crate split

[ArthurZucker]

The README still described the pre-split library. This rewrites it.

What's in it

• Headline on the global set. "Fastest on the world's text, and the only fast one that runs every model." The perf section opens with the 20-cell all-scripts head-to-head against gigatoken (881 vs 590 MB/s, 15/20 cells won) rather than a Latin subset.
• A section that names where we lose. english 0.87×, code 0.82×, and gigatoken leading tokbench's 10-engine ranking at 238.6 vs our 170.1 MB/s — with the explanation that this subset is the intersection of what ten engines support, so engines declining non-Latin shrink it toward English. The coverage asymmetry is stated in gigatoken's favour too: they are byte-exact on all 145 cells they accept, we attempt 203 of 210 and are exact on 174.
• Subcrate map, seven <details> blocks, each answering why is this its own crate rather than just what it contains.
• Bindings matrix — Rust / Python / Node shipping, C & C++ planned, Ruby community.
• Hardware section. The claim is that no SIMD is required: every kernel has an always-compiled portable path that doubles as its byte-exact test oracle. A <details> table lists the four hardware-adapted ops and their fallbacks.
• Footprint, 332799 bytes gzipped, reproducible with make slim-size.

Two figures hotlink from tokbench (figs/03-headtohead.png, figs/01-overall.png).

Caveats, deliberately on the page

Perf figures are the pipeline #2279 + #2296 snapshot and are labelled as such in a <sub> under the table. bitsplit, the thread pool and the load-time fold have all landed since, so these numbers understate the branch — re-running the matrix is follow-up work, not a blocker for replacing a README that describes a library we no longer ship.

Three things found while writing, not fixed here

1. bitsplit's Cargo.toml description advertises Whitespace and Bert grammars that have no modules in models/mod.rs. The README says six (gpt2, cl100k, o200k, tekken, deepseek, kimi); the crate metadata still says eight.
2. tokbench's GitHub Pages site is not deployed, so its own README's huggingface.github.io/tokbench/dashboard.html link 404s. This README links the repo instead. Enabling Pages would be worth it — that dashboard is the best artifact there.
3. The size badge is a static shields URL, so the number needs a manual bump. A CI job to auto-update it was built and then dropped as not worth the machinery for a number that moves a few KB a year.

The branch name mentions a size badge for that reason; the badge automation is not in this PR.