atperson jetstream full-speed ingestion

Live Jetstream ingestion ran at a fraction of the achievable rate: a 200k-event catch-up session observed 5 frames in 60 seconds. Three transport changes close that gap.

Changes

• Drain inside fetch_batch. A WOULD_BLOCK from Wolfram while connected and idle no longer returns to the caller — the batch keeps polling until the feed goes quiet, the budget is spent, or reconnect backoff is advertised. The CLI loop drops its 25ms inter-batch sleep and only sleeps on advertised backoff.
• Cross-batch max-events accounting. The CLI loop applied max-events per batch, so an event-bounded run never terminated. Frames are now counted cumulatively across batches.
• Dictionary-compressed binary frames, on by default. The official Jetstream zstd dictionary is fetched via the public network.bsky.jetstream.getZstdDictionary query on the endpoint's service host and passed to Wolfram's connect options. ATPERSON_JETSTREAM_COMPRESS=0 opts out; fetch failure or a build without libzstd degrades to uncompressed JSON instead of failing the run.
• --kinds filter. Operator-selected event-kind predicate for the v2 subscribeEvents endpoint (--kinds <file> or ATPERSON_JETSTREAM_KINDS_FILE), validated against the four Jetstream v2 kinds, max four unique values. Default remains no predicate, preserving #sync/#identity/#account protocol evidence.

Verification

Both build configs green: core 54/54, network 60/60. New offline tests cover the kinds loader (parse/dedup, unknown kind, limit) and client constructor validation (kinds > 4 rejects; dictionary without libzstd rejects, with it succeeds).

PR: ewanc26/atperson#138