Raise default chunk timeout 45s → 180s so healthy slow generations aren't interrupted #38

Summary

Silence-based stall detection misfires on healthy but slow generations.
DEFAULT_CHUNK_TIMEOUT_MS is 45_000 and gracePeriodMs defaults to 3_000, so a
busy session that emits no events for 48 seconds gets a synthetic "continue"
(the README: "If no events arrive for 48 seconds (chunkTimeoutMs +
gracePeriodMs), the plugin sends "continue" with exponential backoff").

On models with long thinking blocks or slow streaming, 48 seconds is well inside
a healthy generation. Observed locally: a 27B with thinking enabled regularly
goes silent past 60s on one shared GPU, so the plugin interrupts a session that
is making progress and discards its in-flight turn. The v2 port (PR #37) shipped
the same 45s default, so it should be fixed in both for consistency.

Proposed change

- const DEFAULT_CHUNK_TIMEOUT_MS = 45_000
+ const DEFAULT_CHUNK_TIMEOUT_MS = 180_000

plus updating the README prose to match (48 seconds → ~3 minutes).

Notes

• chunkTimeoutMs stays a plugin option (options?.chunkTimeoutMs ?? DEFAULT_CHUNK_TIMEOUT_MS, src/index.ts L493) — anyone who explicitly set a
  value is unaffected; users who want aggressive recovery can set 45_000 or
  busyStallStrategy: "abort".
• Recovery is deferred, not lost: the "continue" still fires with the existing
  exponential backoff and 3-retry ceiling once a session is genuinely silent
  past the window.
• If you'd rather keep the default, a doc-only note calling out the option would
  cover the same ground.

Happy to open a PR with this exact change if you'd like — it's two lines.