fix: don't KeyError on statistics when begin() ran on another thread (#507)

[Mukller]

Fixes #507

Root cause

statistics is a lazily initialized thread-local dict (self._local). begin() populates the four keys (start_time, attempt_number, idle_for, delay_since_first_attempt) on whichever thread starts the call. Under Temporal, the workflow can be replayed so that begin() runs on one thread while the retry loop (iter() → next_action()) is re-executed on another. The second thread's _local.statistics starts as an empty dict, and:

self.statistics["idle_for"] += sleep   # KeyError: 'idle_for'

This matches the reporter's traceback exactly and their suspicion ("a different thread is calling next_action than the one that initializes").

I first tried initializing the property with a full default key set, but that breaks the documented contract that statistics "will be empty when the controller has never been ran" (and an existing pickle test asserts restored.statistics == {}), so this PR keeps the empty-until-begin semantics and makes the write path defensive instead.

Changes

• next_action() now reads with defaults before incrementing:

  stats = self.statistics
  stats["idle_for"] = stats.get("idle_for", 0) + sleep
  stats["attempt_number"] = stats.get("attempt_number", 1) + 1

  A fresh per-thread view therefore starts counting from attempt 1 — the same values a normal run would produce at that point.
• Regression test TestStatisticsKeys::test_iter_on_fresh_thread_after_begin_elsewhere: calls begin() on the main thread, then drives the public iter() from a worker thread with a failed outcome.

Testing

• New test fails on unmodified source with exactly the reported KeyError: 'idle_for'; passes on this branch.
• Full suite: pytest tests/test_tenacity.py -q → 142 passed, 15 subtests passed (CPython 3.13 / Windows).
• ruff check / ruff format --check clean; reno note added under releasenotes/notes/.

[Mukller]

Independently verified this on Windows/CPython 3.13:

• Reverted only the __init__.py hunk and ran the new test against unmodified source - it fails exactly as reported (KeyError path), passes with the fix.
• Full suite green here too: 142 passed + 15 subtests.
• Read-path audit: lines with direct statistics[...] reads are only the two fixed in next_action; the delay_since_first_attempt write further down is a plain dict assignment, which is safe on an empty per-thread dict. So the fix covers all read-modify-write sites.

One thing to fix before merge: the reno note says "Initialize the per-thread statistics dict with the full key set instead of an empty dict", but that is the approach this PR deliberately rejected (it would break the documented empty-until-begin contract and the existing pickle assertion). The actual change keeps the empty default and makes next_action read with defaults. As written, the release notes describe a behavior that does not ship.

Suggested replacement for the note body:

fixes:
  - |
    Don't raise KeyError in ``next_action`` when ``begin()`` ran on a different
    thread than the retry loop (for example under Temporal's replay worker).
    Each thread's lazily created ``statistics`` view now starts counting from
    defaults instead of assuming the keys already exist. See issue #507.

Everything else looks good to me.