plpgsql: enable catching retryable errors in exception blocks

It is possible to catch and handle arbitrary errors within a routine using a PLpgSQL exception block. Once an error is caught, any changes to database state that occurred within the exception block are rolled back, although the most recent values for the PLpgSQL variables are kept. The rollback behavior for exception handling is implemented in https://github.com/cockroachdb/cockroach/pull/110998, but for 23.2 we will disable catching Transaction Retry (40001, 40003) errors because they leave the transaction in a poisoned state that prevents rolling back to a previously created savepoint.

This issue tracks adding support for catching Transaction Retry errors. The read-committed retry loop is able to revert the transaction to a state where savepoint rollback is valid in some cases using [txn.PrepareForPartialRetry](https://github.com/cockroachdb/cockroach/blob/af05d816a1b0b060947114b22d400862b89383c5/pkg/kv/txn.go#L1130-L1135). We will likely do something similar for routine exception handling. Note that some errors can leave the transaction in a state where it must be restarted completely; we will have to find a user-friendly way to handle these cases.

Jira issue: CRDB-31899