Fail fast on rejected keys and unknown models

[rejojer]

llm_completion and llm_acompletion caught every exception the same way:
retry ten times a second apart, then return "". A rejected key or a model
name that does not exist fails identically on all ten, so a run with a bad key
took half a minute to produce a document of empty summaries and exit 0.

They now re-raise on 401, 403 and 404. Those are misconfiguration: every later
call fails the same way, and no caller can do anything useful with the empty
string. Everything else keeps the existing retry, including an unknown status,
which is a transport failure.

400 is deliberately not in the set. It also carries context_length_exceeded,
which is a per-prompt failure on one oversized node rather than a broken setup,
and callers absorb that today. Promoting it would turn one empty summary into a
failed run.

tree_optimize.expand caught the new exception and logged it per node, so the
same misconfiguration went quiet on the --optimize path. It re-raises too.

Measured with an invalid key: a default run now stops in 7s and --optimize in
2.4s, both naming the 401, where before both finished with empty summaries.