Align README, header and installation guidance with earlyon v0.3.0

[sohams25]

Documentation-only alignment of the public face with what v0.3.0 actually is. No product behavior, package version, tag or GitHub Release is changed.

What changed

• Stale wording removed: no more production-ready anywhere user-facing; "wrap anything", "take both savings", "earlyon is the pip install version", and the v0.2-era numeric claims are gone.
• Truthful installation: the primary command is pip install git+https://github.com/sohams25/earlyon@v0.3.0; PyPI appears only as a clearly labelled pending note (the package is not on PyPI yet — trusted publisher not configured).
• Header/banner: subtitle now "Train, calibrate, and benchmark early-exit vision models"; SVGs are labelled illustrative, the "12%" figures and the deprecated computation_used field are removed, alt text corrected. Same visual identity, edited in the SVG sources.
• Per-head calibration + explicit enablement documented: independent temperature per exit and final head, threshold search after fitting, split discipline, explicit disable that cannot fire through a sentinel; small pipeline flow diagram.
• Benchmark evidence updated, negatives preserved: one bounded seeded CIFAR-10 run (1.10× throughput, worse median latency, competitive static MobileNetV2, noise 2.05× labelled synthetic best-case). Legacy v0.2 tables/appendix removed from the README (still quarantined in docs/benchmarks.json).
• Erratum (#6): running the README example exposed that the FLOPs reuse detector false-positives on torchvision ResNet's shared zero-FLOP ReLUs, so the run's estimated fraction (0.88) came from the flagged low-confidence uniform fallback; fvcore attribution gives ~0.92 (~8% estimated saving, not 12%). Documented in the README table footnote, CUDA_EVIDENCE.md erratum, and appended to the evidence JSON without altering recorded fields. Core fix deferred to a future patch per release policy — measured columns are unaffected.
• Static smaller-model comparison: new "Why not just deploy a smaller model?" section.
• Execution modes separated: eager routed vs all-exits ONNX (not conditional execution) vs narrow staged runtime, stated before the deep docs.
• Jetson/TensorRT status: procedures documented, no measurements published, no TensorRT claim, laptop-GPU results don't predict Jetson.
• Checkpoint v2 + migration: what the format records, deterministic v1 migration verified against a genuine prior-version fixture, backup advice, migration-guide link.
• Roadmap: completed v0.3 work removed; explicit feature freeze; narrow evidence-driven items only.

Verification

• All 5 README Python blocks executed against v0.3.0.
• 270 tests, 95% coverage, mypy strict, ruff/black/isort clean, build + twine check PASSED, installed-wheel smoke SMOKE OK, relative-link check clean, secret scan clean.