test: bound the HTTP-date Retry-After delay by the call's own clock window

[lesnik512]

Fixes a flaky test: test_retry_after_http_date_overrides_backoff failed on #156's floors (3.11) job with assert 2.0 <= 1.999698.

Cause

The test built a date "now + 3s" and asserted a delay of 2 to 4 seconds. An HTTP-date has whole-second precision, so formatting drops the sub-second part and the delay comes out as 3 - frac(now) - elapsed. With frac(now) = 0.9997, truncation alone leaves 2.0003, and 0.6 ms of execution then fell below 2.0. A cold first client.get (lazy imports) can also stall for well over a second, which breaks any fixed tolerance.

Fix

The delay retry sleeps is exactly parsed - datetime.now(), taken at some instant during the call. So the test now records before and after around client.get, aligns the date to a whole second, and asserts future - after <= delay <= future - before. That holds however long the call takes, with no tolerance constant. No production code changes.

Verification

The failing case reproduces off-CI: at now = …:00.9997 the old construction gives 2.0003 before elapsed time. The new test passed 30 of 30 consecutive runs locally. just lint-ci passes, and just test-ci passes 1001 tests at 100% coverage.