Fix local usage accounting for resumed sessions and duplicate requests

Fixes #616

Problem and behavior

A resumed rollout can start with the logical session's historical total_token_usage. The local scanner currently subtracts zero for each new file, charging that history to the continuation date. On the reported dataset, the existing algorithm produces 199,153,495 tokens for a day containing 29,044,867 tokens in unique request records.

This updates the shared local usage core used by both the app and daemon:

• Read token_usage_record.payload.usage and deduplicate response_id across files and session roots.
• Treat request records as authoritative in files containing that format. Buffer legacy counters until the file's format is known, so preceding or delayed display events cannot duplicate a request.
• Keep legacy-only files supported: prefer last_token_usage, suppress repeated totals, establish a baseline for the first cumulative-only snapshot and after counter resets, and deduplicate copied events using session ID, timestamp and usage.
• Count input plus output, keep cached input within input, and avoid adding reasoning output or a stale total_tokens field again.

No IPC payload changes or new dependencies are required.

Compatibility boundary

The legacy fallback deliberately does not infer consumption from an initial cumulative-only snapshot, including the first total after standalone last-usage events. That snapshot can contain inherited history; only subsequent increments are attributable. Request records are authoritative at the file level when present. These remain local request-usage statistics, not provider invoice or account-quota data.

Validation

• cargo test --lib shared::local_usage_core::tests --locked: 16 passed, including seven new regressions covering continuations, copied records, cross-root request deduplication, mixed display/request ordering, cached/reasoning tokens, counter resets, aliases and workspace filtering.
• npm run typecheck: passed.
• cargo check --locked: passed (existing upstream warnings remain).
• rustfmt --edition 2021 --check src-tauri/src/shared/local_usage_core.rs and git diff --check: passed.
• A temporary harness executing the modified Rust scanner against sanitized local usage events at the issue's cutoff produced input 28,898,641, cached input 26,920,192, output 146,226, total 29,044,867, matching the independent deduplicated request calculation. This is scanner verification, not live application UI or billing verification. Personal rollout records are not included in the PR.