[ML] Investigate / fix degraded anomaly detection performance on specific QA test sets

The [CStringStore removal](https://github.com/elastic/ml-cpp/issues/2130) has caused degraded performance on a few QA test sets (next to improved / neutral performance on many others, so this doesn't warrant reverting). Investigate what causes this and, if possible, fix it.

The affected test sets are:
- [nginx-remote-ip-request-rate](https://github.com/elastic/qaf-tests/blob/e3ffbfa9b85a7be6c13248c09ef7dece3c22ec8d/resources/test_data/scenario/anomaly_detection/job_nginx_elasticco.yml#L114-L185)
- [nginx_remote_ip_url_count](https://github.com/elastic/qaf-tests/blob/e3ffbfa9b85a7be6c13248c09ef7dece3c22ec8d/resources/test_data/scenario/anomaly_detection/job_nginx_elasticco.yml#L195-L268)
- [iowa_store_liquor_sales_pop](https://github.com/elastic/qaf-tests/blob/e3ffbfa9b85a7be6c13248c09ef7dece3c22ec8d/resources/test_data/scenario/anomaly_detection/job_iowa_liquor.yml#L160-L243)

See also: [this QA issue](https://github.com/elastic/appex-qa-team/issues/156).

Note: no slowdown has been observed locally on MacOS, so this probably needs investigation on Linux or Windows.