common : reject non-positive batch size (#28525)

[goodruyas]

Overview

Currently, if you pass --batch-size 0 or a negative number to llama-server, it triggers a GGML assertion crash deep inside context initialization – not a great user experience. This PR fixes that by:

• Adding validation in the CLI parser (common/arg.cpp) to reject non-positive values outright, with a clear error message.

• Strengthening the check inside llama_init_from_model so it now explicitly rejects n_batch == 0, instead of only complaining when both n_batch and n_ubatch are zero.

• Adding a regression test (test-arg-parser).

• Fixes Eval bug: --batch-size 0 reaches an uncaught GGML assertion #28525

Additional information

• --ubatch-size 0 is left unchanged because it already means "follow n_batch" in llama_init_from_model.

Test Plan

• Build: CPU backend, commit 304665f, GCC 16.2.1, Arch Linux x86_64

• ctest --test-dir build -R test-arg-parser --output-on-failure: 1/1 passed

• Manual: llama-server --batch-size 0 exits with a clean argument error; no GGML_ASSERT.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES – I used AI assistance for brainstorming and to sketch out the initial validation logic and test boilerplate. However, I personally reviewed every single line, fully understand the changes, and can explain the reasoning behind each modification to reviewers. The core decisions – where to put the checks, the exact error wording, and test coverage – were made by me. This description and any future comments are my own, not AI-generated.

[goodruyas]

Thanks for taking a look. I noticed this PR was closed without comments — could you point me to the preferred direction for fixing #28525? If someone is already handling it, I'll leave it to them. I'm new to this repo and want to learn the right way to contribute