Let a managed server finish coming up before giving up on it

[tamnd]

Part of #56. The managed container path was broken in two places and neither failure was the engine's.

First, the readiness wait dials the port, and a server binds its port well before it speaks its own protocol on one. For neo4j that gap is tens of seconds, so the first Ping got an EOF and the run ended with "neo4j: unreachable at bolt://127.0.0.1:32769 (is the server up?)" against a server that was up and starting. A session against a server this run launched now retries for up to ninety seconds. A server the operator supplied still gets one attempt, since that one is either up or not and waiting a minute and a half to say so helps nobody.

Second, the load resolves an import directory by asking the server and then falling back to local-install heuristics. On this machine the fallback found homebrew's import directory, wrote the CSV there, and asked a server inside a container to read it, which ended the load with Couldn't load the external resource at: file:///gb-3991357500.csv. The directory is now proved rather than assumed: one row is written carrying the file's own name and read back through the server, and a directory the server cannot see leaves importDir empty so the load falls back to UNWIND batching. That also covers the case nobody had hit yet, a local install that is not the server on the other end of the connection.

linkbench smoke now runs end to end against a managed neo4j container, all ten queries, and it is the first run to report what that server costs:

Class / Query   p50      p99
point-read      9.43ms   30.57ms
traversal       10.51ms  44.02ms
write           15.38ms  65.50ms

memory peak (sampled)    1.3 GiB
memory steady (sampled)  1.3 GiB
dataset on disk          552.8 KiB
store after load         516.0 MiB
store growth             3.1 KiB

The peak rss row for that same run reads 22.2 MiB, which is the Bolt driver in this process. That is the number the sampled rows replace.

go test ./... green, and the retry has unit tests over a stub engine so it does not need a container to be checked.