Add pre-flight rate limiting with token accounting

[TTK95]

Extends the existing per-provider sliding-window rate limiter with token
windows and a pre-flight check() that blocks requests before they hit
the wire. Stops RWTH (and other) rate-limited providers from surfacing
generic "An internal error occurred" messages after a 429 — the limit
is learned on the first 429, persisted to opencode.json, and on
subsequent runs the pre-flight gate throws a RateLimitError with a
friendly "retry in Ns" message that the retry layer honours via resetAt.

• rate-limit.ts: token windows (minute + day), check() gate, recordUsage,
  estimateRequestTokens, configure() to seed from opencode.json options,
  onRateLimitError persists token limits too
• provider.ts fetch wrapper: pre-flight check, RateLimit.configure from
  options.rateLimit, tick with token estimate
• session/processor.ts: call RateLimit.recordUsage on finish-step so
  the pending estimate is replaced with actual usage
• session/retry.ts: RateLimitError is retryable; delay() honours resetAt
• provider/error.ts: friendly message for 429 responses; mark retryable
• session/message-v2.ts: RateLimitError NamedError, converts to APIError
  shape in fromError so the SDK type surface is unchanged
• config/provider.ts: tokensPerMinute / tokensPerDay schema fields
• cli/cmd/stats.ts: RATE LIMITS section reads persisted limits

https://claude.ai/code/session_01S3gS4AgfQQBNHvgZqkuiRa