feat(dictionary): store full glossary and rank prompt bias

[NathanSkene]

Summary

• store the full custom dictionary as structured entries instead of treating it as a flat string[]
• make SQLite the authoritative dictionary source and migrate legacy custom_dictionary rows into dictionary_entries
• rank and trim dictionary hints per provider budget instead of blindly joining and truncating terms
• expose a glossary import path in the Dictionary UI and show stored vs selected counts for the current transcription mode
• keep reasoning prompts aligned with the same ranked dictionary-selection logic

Why

The current dictionary implementation stores a large list but only really uses a short comma-separated prompt string at transcription time. This change keeps the full vocabulary in-app while sending only the highest-value subset for the active provider/runtime budget.

Verification

• npm run build:renderer
• confirmed the clean branch builds on top of upstream/main
• validated the feature locally on the installed app before splitting this PR out

Notes

• npm run typecheck currently fails on upstream/main because of an unrelated SettingsPage.tsx type issue; this PR does not touch that file.
• This PR is intentionally separated from the overlay/audio work so the dictionary architecture can be reviewed and merged on its own.