fix(langfuse): raise clickhouse nofile limit and bound system log retention

[kokhlo]

Summary

Fixes the "Too many open files" (errno 24) degradation of the bundled langfuse-clickhouse container described in #402: under sustained trace ingestion, background merges over the unbounded system.trace_log exhaust the Docker default nofile=4096, the failing merge is hot-retried forever at ~100% CPU per executor thread, and trace-detail queries return HTTP 500.

Changes

1. docker-compose-langfuse.yml — ulimits: nofile 262144/262144 on the clickhouse service. This is the root-cause fix (ClickHouse docs recommend ≥262k). Mirrors what docker-compose-observability.yml already sets for its clickhouse; works in rootless Docker. Adds a bind mount for a config.d override.
2. langfuse/clickhouse/system-logs.xml (new) — bounded retention for system log tables: TTL 3 days for trace_log/text_log, 7 days for metric_log/query_log/part_log/processors_profile_log/asynchronous_metric_log; logger level warning to stop the debug flood. Without TTLs the giant merges keep being re-scheduled even after the FD limit is raised.

Validation

• docker compose -f docker-compose-langfuse.yml config parses clean; YAML/XML validated.
• The nofile=262144 value and the TTL set were validated live by the reporter in [Bug]: langfuse-clickhouse — "Too many open files" (errno 24) on queries, get_mempolicy log flood, and MergeMutationsExecutor threads spinning at ~100% CPU with no pending mutations #402 (two independent deployments, rootless Docker included) — full diagnostics in the issue.
• Note for existing deployments upgrading in place: on startup 24.10.x renames pre-existing system log tables to <name>_0 (no TTL). The legacy _0 tables need a one-shot ALTER TABLE ... MODIFY TTL or the old data stays on disk — documented in the issue thread.

Closes #402