fix: match reference first-chunk semantics in Wan VAE temporal upsample

[olyasir]

Problem

In Resample::forward (upsample3d, src/wan.hpp), the first latent-frame chunk skips time_conv entirely and never seeds the temporal feat cache, so the second chunk runs time_conv against zero padding instead of chunk-0 context.

The reference implementation (Wan vae2_2.py, "Rep" cache semantics) instead:

• chunk 0: runs time_conv with causal zero padding, doubles the frames, trims the first one
• every chunk (including chunk 0): stores the input tail in feat_cache for the next chunk

Measured effect on Wan2.2 VAE decode vs PyTorch (5-frame 480×832 video, CPU backend, f16 weights): cosine 0.9959 / 27.2 dB PSNR. The output is visually near-identical, so this never shows up in eyeball testing — only in numeric parity. The deviation is concentrated in the first frames of a decode, which matters for any autoregressive/streaming use where early frames feed later context.

Note: upstream leejet/stable-diffusion.cpp has the same logic (src/model/vae/wan_vae.hpp), so this is worth an upstream PR too.

Fix

Mirror the reference semantics exactly (see inline comment referencing vae2_2.py Resample.forward). Encode and TAEHV paths are untouched.

Verification

Measured with an out-of-tree parity harness: it loads the VAE/TAEHV file, runs the real engine graph (CPU) on a raw tensor dumped from the Wan reference implementation, and compares outputs directly — no diffusion model involved. Test input: 5-frame 480×832 video → 2 latent frames.

Results with Wan2.2 VAE + taew2_2 f16 GGUFs (CPU, vs fp32 PyTorch):

Path	cosine	PSNR
VAE decode (before fix)	0.9959	27.2 dB	FAIL
VAE decode (after fix)	1.000000	79.0 dB	PASS
VAE encode	0.999998	89.7 dB	PASS (unchanged)
taew2_2 decode	0.99955	37.3 dB	PASS (unchanged)

[aegioscy]

can you check if this thing even exists before trying to access time_conv.

[olyasir]

Done in e17495a — lookup via find() with GGML_ASSERT on both the map entry and the cast result. Re-ran the parity check after the change: decode still cosine 1.00000 vs the PyTorch reference.