feat(install): check free storage before pulling an LLM model

[mukeshkumarcharak]

Ollama models are 1–20+ GB; on a phone that is often more than the free space, so a blind ollama pull wastes a long download and then fails. Gate the pull on available disk.

• Add _detect_free_disk_mb (df -Pk on the Ollama models dir, walking up to the nearest existing parent), _model_size_mb (per-parameter-size estimate, deliberately a little high), and _largest_model_that_fits.
• In select_and_pull_llm_model: if the chosen model is already on disk, pull needs no space. Otherwise, if it doesn't fit, downgrade to the largest model that does; if nothing fits, reuse an already-pulled model or skip the download with a clear "free up space, then: ollama pull gemma3:1b" message instead of failing mid-download.
• verify()'s "model not downloaded" report now notes when the cause is low storage and adjusts the fix hint accordingly.