fix(compression): deduplicate applied diagnostic log per unique compressed output

Previously the compression service emitted 'Tool output compression
evaluated' (INFO) once per applied tool message in every request,
including repeats of the same compressed output carried across longer
histories. Add an in-memory dedupe cache keyed by
(tool_call_id, original_sha256, compressed_sha256, selected_rule)
so applied logs emit only the first time a given compression result
is produced. Fail-open logs remain unsuppressed.

Also fix stale import in test_responses_stream_legacy.py
(responses_stream_legacy -> responses_stream_coercion).

Author: Mateusz