fix(stats): count forked sessions' copied messages once

[slkiser]

Summary

• When OpenCode forks a session, it copies the parent's finished messages into the fork with new ids but the same times, model, tokens and cost. The all-history, time-window and multi-session readers counted each copy again, so forking a $10 session without using the fork showed $20 in token/cost totals (/tokens_*), Cursor usage and local-estimate quotas.
• dropForkCopies in src/lib/opencode-storage.ts now counts each finished fingerprint (created/completed time, provider, model, tokens, cost) as many times as the session that has it most often, and drops the extra copies. Identical rows inside one session are kept; unfinished rows are never merged.
• A single session's own read (/tokens_session on a fork) is unchanged and still shows its copied history.
• Known limit: two unrelated sessions with exactly the same created/completed millisecond, model, tokens and cost are counted once. Fork links aren't used because OpenCode 1 forks migrated into OpenCode 2's database have none.

Linked Issue

Refs #306 (reported on 4.x; the 4.x backport is a separate PR into release/4.x).

OpenCode Validation

• Current production released OpenCode version tested: v2.0.20
• Why this version is relevant to the fix: its session.forked projector copies the parent's finished session_message rows into the fork; the local package build was run in an isolated OpenCode 2.0.20 and a session was forked.

Before-and-after evidence

Before: not captured live. By the issue's steps, the fork's copied messages are counted again (here: 6 messages, about 22.9K tokens).

After (/tokens_today, same window, before and after forking "Quick check-in" with Full session and sending nothing in the fork):

Before fork:  Messages 4 | Sessions 2 | Tokens 15.7K | Cost $0.00
After fork:   Messages 4 | Sessions 2 | Tokens 15.7K | Cost $0.00

Surface checks:

• /tokens_today command dialog (TUI): checked live, totals unchanged after the fork.
• TUI sidebar, toast, compact line below the message input: not applicable; they show provider quota, not local token history.
• Web output: not separately tested (same reader).

Quality Checklist

• I ran pnpm verify
• This change is focused and avoids unrelated behavior changes
• I updated or added tests when behavior changed
• I updated docs when user-facing workflow, command, or config behavior changed (no docs describe how totals are counted)
• For provider changes, I followed Provider Changes, or this does not apply