feat: real-time Token Savings panel and tiktoken cross-check

Surfaces the context_savings metadata introduced in 2.3.4 as a boxed
multi-line panel on the CLI, so humans (not just MCP-side LLM agents)
see how much context the graph kept out of the window.

- `format_context_savings_panel()` in context_savings.py renders the
  panel with a per-category breakdown (Functions / Tests / Risk /
  Other) that sums exactly to the graph response size — no padding,
  no rounding.
- `--verify` flag on `detect-changes` and `update --brief` cross-checks
  the displayed numbers against tiktoken's cl100k_base tokenizer
  (the GPT-4 family). Calibration in docs/REPRODUCING.md shows the
  chars/4 approximation stays within +0.5% of real tokens in
  aggregate across 222 mixed-language source files.
- `update --brief` is new: incremental update plus the same risk
  panel in one command. Distinct from `detect-changes --brief`
  (read-only against the existing graph).
- `code-review-graph embed` CLI subcommand for explicit shell-level
  access to embedding generation. Previously only reachable via MCP.
- `_get_version()` now falls back to package `__version__` when
  importlib.metadata can't find the dist-info (iCloud / OneDrive
  edge case).

Tests updated to assert the panel structure instead of the old
one-line "Estimated context saved:" format.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>