Latent-space chaining seams for chunked video continuation

[TroyHernandez]

Chained multi-chunk generation today round-trips every join through pixels: decode chunk N, H.264-encode it, re-read its tail frames, VAE re-encode them into conditioning latents for chunk N+1 — plus a full onload/offload cycle of every component per chunk. This PR adds the seams to keep a chain in latent space end to end and to hold components resident across back-to-back calls.

New surface

• condition_latents= on txt2vid_ltx2(): an already-encoded conditioning prefix [1, 128, k, H/32, W/32], bypassing the read/preprocess/VAE-encode block entirely. Mutually exclusive with image/condition_video; geometry-validated.
• ltx23_tail_latents() (exported): slices the trailing k latent frames of a result into condition_latents form. Results now carry latent_shape to make that self-contained. The docs flag the semantics caveat: sliced continuation latents represent 8 pixel frames each, a fresh tail encode represents 1 + 8(k−1) with a first-frame-form leading latent — A/B on real content before trusting latent-only joins.
• resident=: component names to keep on the compute device after their phase. Onload is now idempotent (device check, tryCatch-guarded so anything odd falls back to the old path), offload and the dequant-buffer release skip resident components. A chunk loop pays one transformer onload per track instead of one per chunk.
• trim_frames=: drop leading pixel frames (the conditioning-head overlap) from the decoded video and the saved file, in memory — replacing a lossy subclip re-encode downstream. The returned latents keep the full sequence, since tail slicing needs it.

Head trim stays in pixel space by design: the causal VAE decodes latent frame 1 to a single pixel frame and later frames to 8 each, so dropping leading latent frames before decode yields the wrong frame count (7 short) and pushes a continuation-form latent through the decoder's first-frame path.

Tests

Extends the tiny-pipeline CPU fixture: condition_latents reproduces the condition_video path bit-for-bit when fed the same encoded tail; tail slicing matches a manual unpack+narrow; a sliced tail feeds back end to end; trim_frames output equals the untrimmed video minus its head, with latents and latent_shape intact; guardrails for the three-way source exclusivity, bad geometry, bad trim, and unknown resident names. Full suite: 923 green (CPU).

GPU validation still owed (deferred while the card is busy): resident-transformer chain timing, and the sliced-tail vs fresh-encoded-tail A/B on identity/lip-sync stability.