aws/cloudtrail: clean up transient json and _conf fields in on_failure handler

When the ingest pipeline fails at any processor, the top-level on_failure
handler catches the error but does not remove the transient `json` and `_conf`
fields. These fields are normally removed at the end of a successful pipeline
run, but on failure they leak into the indexed document.

Because CloudTrail aggregates events from many AWS services with inconsistent
schemas (e.g. `json.requestParameters.filter` is a string from IAM Identity
Center SCIM but an object from AWS Health/EC2), the leaked `json.*` fields
create dynamic mappings on the first failure, then reject subsequent documents
with a different type for the same path, causing document_parsing_exception
errors and filling up the Logstash DLQ.

Adds a remove processor for `json` and `_conf` to the on_failure handler,
mirroring the cleanup already done in the normal pipeline flow.