feat(import): add ccusage import format (#1128)

Adds `--format ccusage` to `tokscale import`, alongside the clawdboard
importer. Part of #888.

`ccusage daily --json` is the export most people already have on disk, and it
usually outlives `~/.claude/projects` after rotation — so it is the common
recovery path for history that can no longer be re-scanned.

Attribution resolves per model rather than per row, since a unified ccusage
day merges every agent that ran into one row and names the participants only
in `metadata.agents`. A model matching no listed agent is surfaced as
`unknown` rather than folded into the first agent, because the leaderboard
rejects `unknown` and a clean-looking wrong import is worse than a visible gap.

Followed by a maintainer commit validating the parser against a real `cc.json`
and against ccusage's documented schemas, which corrected four divergences:
reasoning tokens counted twice (ccusage's `totalTokens` never includes them,
so the buckets now hold the non-reasoning output remainder); reasoning read
only from the flat spelling when the unified report nests it under `metadata`;
an empty day-in-range row minting a phantom `unknown` client with a spurious
rejection warning; and `ccusage-codex`'s `costUSD` spelling importing as $0.
Agent names keeping a CLI suffix (`gemini-cli`) also failed to match their own
models. The row's declared `totalTokens` is now reconciled as a cross-check.

Known limit, documented in `resolve_client`: attribution splits by model
family, not by the client that made the call, so a day listing both `claude`
and `opencode` puts all Claude models on `claude`. Exact splitting needs
ccusage's `--by-agent` rows.

Verified: a real 14-day export round-trips to 267,186,765 tokens / $453.81,
matching the file's own totals exactly. `cargo test --workspace` 3065 passed /
0 failed; clippy `-D warnings` and `cargo fmt --check` clean.

Co-authored-by: Aron Lee <hello@intframe.com>