feat(token-usage): speed-keyed buckets and repricing fields on the wire

Fast and standard usage of one model bill at different rates, so speed
(0 standard, 1 fast) joins the bucket identity: bucket_state is keyed by
(session_id, model, speed, bucket_ts) (schema v4 rebuilds just that
table; entries and cursors survive) and aggregation groups by the
entry's speed, with unmarked entries falling into the standard bucket.

TokenUsage events gain positions 11-19 so the server can recompute a
bucket's cost under a different pricing sheet: speed and whether it was
inferred from configuration, the 1h cache-write split, the long-context
token splits per class (tokens of requests that selected their model's
long-context tier; whole-request selection already applied client-side,
base-tier tokens are the totals minus these), and the transcript-priced
portion of the cost, which is fixed under repricing (its tokens remain
in the totals, so such buckets reprice approximately). The pricing
catalog id rides custom_attributes on catalog-priced buckets.

The bucket fingerprint covers every new emitted value, so any split
changing re-emits the bucket at a higher revision; the catalog id alone
is deliberately excluded (an id change without numeric change means
identical rates). The spec's value table, bucket identity, and pricing
sections are updated accordingly.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>