Validate Nitter response content to prune broken instances

[dreikanter]

Summary

Fixes the Feedjira::NoParserAvailable errors (538+ occurrences) affecting every Nitter-backed feed.

Root cause: the Nitter subsystem validated only HTTP status, never response content. A Nitter instance returning HTTP 200 with a non-RSS body (an HTML rate-limit/error page) passed both the availability check and the loader's status check. The HTML was then handed to FeedjiraProcessor, raising NoParserAvailable. Because the response was 2xx, register_error never fired, so the broken instance was never suspended — it stayed in the pool and corrupted ~1/N of every Nitter feed's pulls. This is why feeds like neural-machine got fresh posts when they drew a healthy instance but errored when they drew a poison one (instance-level, not account-level).

Fix — validate that the response is actually a parseable feed, in both places:

• NitterInstanceAvailabilityChecker#available? now requires the sample RSS to parse, so poison instances are marked disabled on pool refresh.
• NitterLoader now treats unparseable content as an instance error (register_error + raise NitterLoader::Error), so a poison instance that slips through gets suspended after MAX_ERRORS and the failure is correctly attributed and self-heals.
• New FeedContent.parseable? helper, shared by both.

Healthy instances keep serving the live feeds; broken ones get pruned.

Also disables 4 inactive Nitter feeds. Their X source accounts have produced no content for 16 months to 4 years (verified against freefeed cadence and source research): catscafe, extrafabulous, jakelikesonions, nathanwpyle. Each carries a disabling_reason. The live feeds (neural-machine, warandpeas, yesbut, safelyendangered) stay enabled; their 2h refresh interval is already well within Nitter rate limits.

Checklist

• Updated CHANGELOG.md under Unreleased, or change is internal-only