fix(llama-cpp): explain tensor count mismatch

[localai-org-maint-bot]

Description

This PR preserves llama.cpp tensor-count load errors and appends an actionable hint that the GGUF layout may be incompatible with the selected backend or the file may be corrupt. It leaves unrelated load errors unchanged.

Ollama-published monolithic Gemma 4 blobs use Ollama's private compatibility layer to hide embedded projector tensors. LocalAI uses upstream llama.cpp, so its loader sees the extra tensors and rejects that internal blob layout. The model-scoped backend process then exits after failed initialization; LocalAI itself remains available.

Closes #11593

Verification

g++ -std=c++17 backend/cpp/llama-cpp/model_load_error_test.cpp -o /tmp/localai-model-load-error-test
/tmp/localai-model-load-error-test
git diff --check

The standalone regression test covers the exact expected 2131, got 720 diagnostic, a generic tensor-count mismatch, and an unrelated load error.

Notes for Reviewers

The change deliberately does not import Ollama's compatibility patches, pre-parse GGUF metadata, or alter RPC status and backend lifecycle.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable