Fix AIR live log polling correctness

[stevenchen-db]

Summary

• poll Bricklens at a fixed one-second rate without adding another sleep after slow requests
• retain a bounded lookback and deduplicate records so late and out-of-order logs are not dropped or repeated
• wait for consecutive empty terminal polls so recently emitted records can become searchable
• bound existing logs when attaching to an active run while preserving explicit --minutes behavior

This matches the AIR Python CLI behavior in databricks-eng/universe#2611686.

Validation

• ./task lint-q
• focused AIR command tests with -count=1
• AIR logs and run-watch caller tests with -count=1
• GOFLAGS=-buildvcs=false ./task build

[eng-dev-ecosystem-bot]

Integration test report

Commit: 838f580

Run: 35145772100

	Env	✅​pass	🙈​skip	Time
✅​	aws linux	276	17	4:22
✅​	aws windows	278	15	3:33
✅​	azure linux	275	17	4:02
✅​	azure windows	277	15	3:43
✅​	gcp linux	276	17	4:25
✅​	gcp windows	278	15	3:37

Top 3 slowest tests (at least 2 minutes):

duration	env	testname
3:39	azure windows	TestAccept
3:35	gcp windows	TestAccept
3:30	aws windows	TestAccept

[caroline-db]

Initial tail may leak omitted history:

After printing the newest N records, the code moves fromSec back by 30 seconds but remembers only those N records. The next poll fetches the overlap and prints older, omitted records as if they were new.

Example: an active run has lines 1, 2, 3; --tail 2 can print 2, 3, then 1 on the next poll.

This defeats both the line limit and chronological ordering. The Python implementation has the same bug, so the Go port matches it but is not independently correct.

(not from this PR but may be relevant) Go’s --tail 0 diverges from Python: Go accepts it and later prints the backlog, while Python rejects zero.

[caroline-db]

implementation looks good but some tests still need to be updated to reflect the new help text and to fail for --tail 0.