fix(tracking): use weighted savings rate in low_savings_commands and avg_savings_per_command

AVG(savings_pct) over rows produces an unweighted mean: a handful of 0%-savings
passthrough calls dilute a filter that genuinely performs well on high-volume
invocations. Same root cause as PR #891 (get_by_command).

Changes:
- low_savings_commands: AVG(savings_pct) → SUM(saved)*100/SUM(input). A command
  with 1 big call (95%) + 4 small calls (0%) now shows 94.6%, not 19%. Filters
  correctly excluded from the "needs improvement" list shipped in telemetry.
- avg_savings_per_command: inner per-command rate is now weighted (SUM/SUM);
  outer average across command names stays unweighted (intentional, each filter
  counts once). Documented in TELEMETRY.md.
- Two regression tests added (conn.execute direct inserts for DB isolation).
- docs/TELEMETRY.md updated to reflect the weighted formula.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Florian BRUNIAUX <florian@bruniaux.com>