#648 native dense: run the architecture guard only on the first shard of a split

Beschreibung

## What

`NativeDense::load` ran `check_architecture` on **every** shard that carries `general.architecture`. A splitter may copy that key into **all** shards of a split while the `qwen4exp.*` hyperparameters live only in shard 1 — `NativeDense::load` then fails on shard 2 with:

```
strata generate: native dense projections: missing qwen4exp.block_count
```

Seen with a re-split of the UD-Q4_K_XL GGUF (the Huihui abliterated release of Qwen3.8-Flash-Next: `general.architecture` present in shards 2-4, no `qwen4exp.*` there).

## Change

Run the full architecture guard only when the shard is the first of the split (`split.no` absent or `0`); later shards take the existing split-metadata consistency check whether or not they carry the architecture key. Single-file GGUFs are unaffected (they always have `split.no == 0` or none).

## Verification

- Before: the engine refuses the re-split GGUF at load (error above).
- After: loads end to end — pack built with `iq_pack.py --compat-bf16`, resident mode, coherent greedy output from the 125B.
- `strata-gguf <shard1> --check` still passes (the guard on shard 1 is unchanged).

Tested on the HIP backend (gfx1030 / RX 6950 XT, ROCm 7.1).