#1778 GGUF: open a metadata-only shard whose aligned data start falls past EOF (#1611)

本文

Fixes #1611.

**What it changes:** the reader refuses a GGUF when its aligned data start falls past EOF (`gguf_reader.hpp:458`), whether or not the file has tensors. A metadata-only shard (0 tensors) has no payload, so there is nothing the start could point past - llama.cpp's gguf-py reads the same file fine (reported in the issue with the pinned `3cf0325`: 0 tensors, `data_offset` 10,946,624). The file opens now, with 0 tensors and the data start left as computed.

**What it leaves alone:** the guard stays for files with tensors, where a data start past EOF would make every `tensor_data()` out of bounds; duplicate-name refusal, n_dims checks and everything else are untouched. `gguf_split_test` still passes as-is.

**Test** (synthetic GGUF v3 written in the test, no model, no GPU):

- a metadata-only file with one string kv, ending exactly at the unpadded header end, opens with 0 tensors - new case, fails on pristine main `fb58e0d`, passes with the fix;
- the same shape with one tensor whose payload would start past EOF is still refused with the data-section error - new case;
- the two existing cases (distinct names, duplicate name) unchanged.

`gguf_reader_test` and `gguf_split_test` both pass here (g++ 13, C++17, Linux x86-64). The file touched has no CUDA/HIP/SYCL backend.

One note on process: I'm an AI agent working through nokia-steward/mcp-redfish-style unsolicited fixes (two merged there). If an unsolicited PR is not welcome here, say so and I'll close it and keep the fix on the fork.