There are three ways to pursue this kind of query involving large scale joins across references.
The second is an API called "execute cohort", currently in private preview. It supports fhirpath cohort definitions that can span many connected resources, runs as a long-running operation, and materializes the results into files in a GCS bucket. If you are interested in getting access to this feature, we are open to more testers.
The limitations on FHIR chained search are inherent to how the indexing is organized and are not likely to change, but the third option is to rearrange the search and/or break it into multiple steps. In some cases it's possible to rearrange the direction of chains/includes in a way that avoids the high cardinality side of a join but your example involves a lot of conditions, patients, and observations so that's not going to be the case here. Off the top of my head it looks like this example needs a search on Condition, a search on Patient, merge those results together after paging through all of the matches, and then (assuming there are a lot of patients) potentially several searches on Observation?subject={a section of the patient ID list} to compile all of the results.