Strata Manager: descargar versiones anteriores

- Adding an engine argument no longer writes it twice (#226): when the config already had custom arguments (user_args), one click on "Add argument" in the full engine argument list wrote the same option twice. It now adds one copy. A single-use option that is already in the list is updated in place (the button reads "Update item N" and the row is highlighted); repeatable options such as --control-vector and unknown options are still appended.
- Model preparation says what differs from the current settings (#225): when "My models" already has the same model and this preparation differs in model name, weight files, context or image input, the task used to say only that the managed snapshot differed. The task now shows "Needs your choice" instead of "Completed", lists each difference with the current and prepared values (for example context 160K now, 128K this time), and offers Keep current settings, Change settings and Detect again.

- Set GGML_NO_BACKTRACE=1 on engine child processes: matches upstream 38ad0bc. ggml's abort path may fork gdb/lldb and write debugger errors into the engine log and the picture reader's READY pipe. Starting a service through serve.server, running setup.py tasks, reading engine --help, and the upgrade trial --version now set the variable in the child environment only (an existing value is left alone; STRATA_BACKTRACE leaves it unset). Manager's own process environment is unchanged.

- Read every gzip member when extracting official tar.gz packages: an upstream feed may emit concatenated multi-member gzip when packing with multiple threads (not the default yet). GzDecoder only read the first member, so later files failed to extract. Extraction now uses MultiGzDecoder; single-member packages behave as before.
- Do not offer image input when the engine package has no strata-vision: official 0.1.59 SYCL packages (upstream #5716) ship without a picture reader. Setup, model preparation, and image settings no longer default-enable or allow turning on image input, and a short note explains that only text is supported. Such packages are still accepted; packages that include a vision binary are unchanged.