The URLs for addresses are meaningless until the node http addr
is available.

Vendor in github.com/hashicorp/go-safetemp

Shorten docker health timeout

Only run health check if driver moves from undetected to detected

rpc: only attempt NodeRpc for nodes>=0.8

DRY it up

Attempting NodeRpc (or streaming node rpc) for clients that do not
support it causes it to hang indefinitely because while the TCP
connection exists, the client will never respond.

Correctly initialize array length

docs: mention env var changes in upgrade docs

Mention the changes from #3760 in the upgrade docs as applications
expecting underscores will break.

UI: Unlink record relationships before unloading them from the store

It also culls for findAll requests, so it too needs to be
careful about leaving garbage around.

When simply unloading a record, references to the record are
maintained by the internal relationship state of related models.
This causes refetching and duplicate models local to that relationship
state.

Never respond with an empty object from the adapter

In production builds only, this will slip into the local store
as a record with no ID, which makes for all sorts of bad news.

Driver health UX improvements

Improved Client handling of failed RPCs

health check timeout should be 1 minute

only set cached connections upon no error

separate drivers by commas