feature: per-session and per-model token and cost aggregation

Feature Description

The primary deliverable of the release. Today the numbers are per-event: you can see a request,
but not what a session, a model, or a day cost.

The parser already records the full breakdown — parsercommon.TokenUsage carries input,
cache-read, cache-write and output — so what is missing is aggregation, not measurement.

Proposed Solution

• Aggregate per session: input / cache-read / cache-write / output tokens, cost, wall time
• Aggregate per model and per endpoint
• Cache-read and cache-write reported separately — they price very differently and
  conflating them makes every cost number wrong
• Aggregates survive a restart (depends on #901)
• One documented field-name schema, shared by the TUI metrics view, unattended workloads
  and the central collector (#898)

Additional Context

Part of the Cortex v0.9.0 release. Depends on session differentiation — per-session aggregation
is meaningless while two concurrent sessions cannot be told apart. Fix #900 first: that is the
narrower "no token shown at all" bug.