Add kernels.status_history JSONB column to track the datetimes for each status change

[achimnol]

Let's allow users to query when their sessions and kernels enteres which status.
(e.g., when did this session move from PENDING to SCHEDULED, PREPARING to RUNNING, etc. afterwards)

Fortunately, there is no backward status transition in the kernel's lifecycle.
(e.g., PENDING -> SCHEDULED exists but SCHEDULED -> PENDING does not exist)
So we could use a simple JSONB column whose keys are the status names and their values are the timestamps when the row's status is set as the key.

{
  "PENDING": "...",  // would be same to created_at
  "SCHEDULED": "...",
  "PREPARING": "...",
  "RUNNING": "...",
  // ...
  "TERMINATED": "...",  // would be same to terminated_at
}

When the key does not exist or null, it means either the kernel has not been gone into that status or the record is a migrated one from older versions.

Let's use manager.models.utils.sql_json_merge() to incrementally add a new key as follows:

sa.update(kernels).values(..., status_history=sql_json_merge(
    kernels.c.status_history, [], {"<new-status>": "<timestamp>"},
))

For maximum interoperability, the timestamp values could be formatted as UTC-based ISO8601 datetime.

Q. How could we "ensure" all database status change queries to update this json field altogether?

[rapsealk]

This feature would be also helpful to the future work on #395.