Split NCBI outages from paper-level text-unavailable in PmcBiocFetcher

[bobular]

Prompted by a real incident: the whole /research/bionlp/RESTful/* tree returned 503 "no healthy upstream" for hours while the rest of NCBI was fine, but we reported it to users as "this article has no full text" and PmcBiocFetcher logged nothing at all, so the outage was invisible server-side.

Transport failures / 5xx / 429 now raise UpstreamUnavailableException -> terminal upstream-unavailable (not cached, not publishable), kept distinct from text-unavailable since the two call for opposite user actions. Added log4j logging (error/warn/info per outcome, including the response body snippet that identified this as NCBI's problem) and a javadoc note on BIOC_URL_BASE recording the PubTator3 endpoint as a verified fallback source. Replaced the unused HttpClient constructor seam with a narrow BiocHttpExchange interface so every branch of the failure taxonomy is unit-testable without network.

Front-end companion change (new upstream-unavailable status, messaging, recovery buttons) lives in the web-monorepo repo and ships together. VEuPathDB/web-monorepo#1843

---

Screenshot of new behaviour from the accompanying front end PR (old behaviour below)

Old behaviour

[bobular]

Reviewed/understood all the Claude code and tested on my dev site. Log message (wdk.log) for the current problem looks like this:

ERROR - org.apidb.apicommon.service.services.ai.article.PmcBiocFetcher:158 - PMC BioC returned HTTP 503 for PMID 21533217 [https://www.ncbi.nlm.nih.gov/research/bionlp/RESTful/pmcoa.cgi/BioC_json/21533217] after 59 ms; treating as upstream outage. Body: no healthy upstream

Client devtools shows this:

{
	"reason": "the PMC BioC service returned HTTP 503 for PMID 21533217",
	"job_id": "2077e3900a195b7a95d40a38a22a91cc6acd527731e1d1467e9fd412630a53a3",
	"type": "upstream-unavailable"
}

Client browser shows message in main PR description screenshot.