chore: add one-turn internet prep bundle workflow

[rad1092]

Motivation

• Provide a reproducible, one-turn online workflow to collect environment metadata and assemble an offline bundle for later offline installation and model assets archival.
• Make offline artifacts (.offline_bundle/) easy to generate and exclude from source control to avoid accidental commits.
• Stabilize baseline comparisons for CSV analysis by adding a baseline plan and representative fixtures for regression checks.
• Improve numeric parsing so the analyzer correctly recognizes common human-entered numeric formats (thousands separators, currency symbols, percent signs, and negative-by-parentheses).

Description

• Add scripts/prepare_online_bundle.sh which collects Python/pip metadata, builds a local project wheel, attempts to download optional dependency wheels, archives the Ollama installer script if reachable, and emits an offline usage guide at .offline_bundle/OFFLINE_USE.md.
• Add ONLINE_EXECUTION_LOG.md, BASELINE_PLAN.md, and REVIEW_REPORT.md documenting the one-turn run results, baseline criteria, and an execution/analysis report.
• Update .gitignore to exclude .offline_bundle/ artifacts.
• Normalize numeric parsing in bitnet_tools/analysis.py by stripping ,, ₩, $, €, £, % and handling negative values expressed in parentheses in _to_float.
• Add representative CSV fixtures under tests/fixtures/ and extend tests/test_analysis.py with test_to_float_normalizes_currency_comma_percent_parentheses and test_baseline_fixture_summaries_are_stable to assert normalization behavior and baseline summary stability.

Testing

• Ran pytest -q, all tests passed (26 passed).
• New unit tests covering _to_float normalization and fixture summary stability were included and passed under the same test run.

---

Codex Task