[andared]

Refactor with no intended change in behaviour.

• The load token goes through setValue, setNegative, store and shard.set as one nullable *loadToken parameter instead of a variadic tokens ...*loadToken that only ever held zero or one.
• The recheck under the flight-group lock, which was duplicated in core.load, core.acquire and View.load, is now shard.beginLoadUnlessCached. The view passes isType[T] so that a hit of another type still loads.
• The repeated idx := c.shardIndex(key); g, s := c.flights[idx], c.shards[idx] is now flightFor, on both core and View.
• Len and Bytes read only the shard sizes instead of building a full Stats snapshot, which loaded all 15 atomic counters of every shard.
• Comments that restated the code are removed. The comment on shard.set now documents the non-obvious part: a stale token makes the write a silent no-op. Comments in tests explain intent and are left as they are.

Checklist

• make test and make lint pass (go test -race ./..., golangci-lint run: 0 issues)
• Behaviour changes come with a test that fails without the change: n/a, no behaviour change
• Performance claims come with benchmark output: none made; Len/Bytes only drop work
• Public API changes are noted in CHANGELOG.md: no API change; the Len/Bytes internals are noted under Unreleased
• No new dependencies