fix(retry): cap backoff exponent to avoid OverflowError on large attempts

[winklemad]

Fixes #1760.

Problem

anthropic.lib._retry.backoff(attempt, *, cap, base=2.0) returns min(cap, base ** attempt). Because base is the float 2.0, 2.0 ** attempt overflows a C double at attempt >= 1024 and raises OverflowError before min() can clamp it, so the documented cap is never returned:

from anthropic.lib._retry import backoff
backoff(1023, cap=60.0)   # 60.0
backoff(1024, cap=60.0)   # OverflowError: (34, 'Result too large')

The only caller is the environments poller (lib/environments/_poller.py, _backoff = backoff(attempt, cap=60.0)). Its counter resets to 0 on success, so this needs ~1024 consecutive transient failures — but at that point the run-forever poll loop dies with an uncaught OverflowError instead of continuing to retry at the cap.

Fix

Cap the exponent at 1000, mirroring the guard the core client already applies in _base_client.py:

# Also cap retry count to 1000 to avoid any potential overflows with `pow`
nb_retries = min(max_retries - remaining_retries, 1000)

backoff() just missed the same guard. The returned value is unchanged for every attempt — once base ** attempt reaches cap the result is always cap, which happens far below attempt = 1000 — so this only removes the crash.

Tests

Extended the existing test_backoff (tests/lib/environments/test_poller.py) with the overflow-boundary cases (attempt=1024 and attempt=5000, both expecting the 60.0 cap). They raise OverflowError before this change and pass after; the existing cases are unchanged.

Verified locally with ruff check/ruff format clean.