docs: fix metadata, slim the README, widen CI to 3.10-3.13

[TheronRadley]

What this does

Works the "quick wins" and the two documentation items from the roadmap, top to bottom.

Quick wins

• Clone URL — github.com/example/tiny-git → github.com/TheronRadley/tiny-git. Also fixed four project.urls in pyproject.toml that were still pointing at example/.
• Intro & naming — removed the duplicated tiny-git (tiny-git), and added a table explaining the hyphen once: tiny-git is the distribution and the command, tinygit is the importable package, because hyphens are not legal in Python identifiers.
• Live badges — replaced [Tests: 98 Passed (0 Failed)] with the workflow's own badge, and stripped baked-in counts from the prose. Counts now live in docs/testing.md with the command to regenerate them.
• Topics — ⚠️ blocked, needs you. gh repo edit --add-topic returns HTTP 403: Resource not accessible by integration. The token reports admin: true on the repo but the app installation lacks the Administration permission that the topics endpoint requires. Add python, git, version-control, educational under Settings → General → Topics, or grant the app that permission.
• Social preview — image generated and committed to docs/, but GitHub has no API for social previews; it must be uploaded under Settings → General → Social preview.

Docs

• Slim README — invariants table, test breakdown, and benchmarks moved into docs/ behind a documentation index. The README keeps the pitch, install, quick start, and command list.
• Git-compatibility note — new "Not wire-compatible with real Git" section. Verified in src/tinygit/objects.py: Tree.serialize() emits UTF-8 mode sha name\n lines, and .tinygit/index is JSON. Real git reads binary trees and a binary index, so neither tool can read the other's repository.

Packaging & CI

• Python floor — requires-python lowered >=3.12 → >=3.10, classifiers and ruff/mypy targets follow. Matrix is now ubuntu + macos × 3.10–3.13, fail-fast: false.
• Windows — added as a non-blocking job. os.chmod(0o755) only flips the read-only bit on Windows, so the executable-mode round-trip tests cannot pass there. The four specific call sites are named in a workflow comment; promote the job once they skip on platforms without POSIX mode bits.
• .gitattributes — * text=auto eol=lf, so blobs and fixtures are never rewritten on a Windows checkout.
• PEP 639 license — license = "MIT" with setuptools>=77, clearing the build deprecation that would otherwise block a PyPI publish (deadline 2027-02-18).

CHANGELOG

Collapsed the 1.0.0 / 1.1.0 entries into a single 1.0.0. Neither version was ever tagged or published, every feature listed under both is present in the shipped code, and 1.1.0 stays reserved for the roadmap's tag / count-objects / gc work — so this file, pyproject.toml, and the v1.0.0 tag will agree.

Corrections to claims that were wrong

Claim	Reality
fsck sample output showed ✓ 3 objects checked	The quick start immediately above it produces 6 objects (2 blobs + 2 trees + 2 commits). Output replaced with a real captured run.
"full CLI workflows … live CLI subprocess execution"	grep -rn subprocess tests/ returns nothing. test_cli_contract.py calls main() in-process. Noted as an open gap in docs/testing.md.
Benchmark table listed status and diff rows	run_benchmarks.py emits status_staged, status_clean, and diff_clean. The table did not match the script that produced it.
"Over 25 dedicated test functions" for merging	27 across six files — roughly right, but now stated without a number that will rot.

Verification

Run locally on Python 3.11.2 (below the old >=3.12 floor):

pytest                     -> 98 passed in 3.96s
ruff check src tests       -> All checks passed!
mypy src                   -> Success: no issues found in 16 source files
mypy --python-version 3.10 -> Success: no issues found in 16 source files
python -m build            -> sdist + wheel, no deprecation warnings

The wheel was installed into a clean venv and the tiny-git console script exercised end to end (init → config → add → commit → fsck → log). Benchmarks were re-run at 100/1,000/2,500/10,000 files; all operations scale between 9.1× and 10.5× for a 10× file increase, i.e. linear.

3.10 is asserted by this PR's CI run, not by a local run — there is no 3.10 interpreter in the sandbox. If the 3.10 leg goes red, the floor needs to move back to >=3.11.

Follow-ups (not in this PR)

• Topics and social preview need repo-settings access.
• Windows: guard the four exec-bit assertions, then promote the job into the blocking matrix.
• A real subprocess CLI test via sys.executable -m tinygit.