feat(server): support transient null overrides on load

[fl0rianr]

Summary

Add transient per-option clearing semantics to POST /v1/load.

An explicitly provided null now means: ignore the saved per-model value for this load only and fall through to the lower server/default layers.

This gives /v1/load three distinct states for recipe options:

• omitted: keep using the saved per-model value
• null: ignore the saved per-model value for this load only
• concrete value: override it for this load

ctx_size: -1 remains the explicit auto-sizing value and is not treated as a clear.

Why

This lets clients clear individual saved options without mutating recipe_options.json.

For example, a client can temporarily clear saved llamacpp_args without also losing unrelated saved settings. This is more precise than a global ignore_saved_options flag.

Tests

Added coverage for:

• saved args → transient null clear
• another saved option remains active
• merge_args behavior
• explicit ctx_size: -1

[bitgamma]

there is one edge case that is probably worth addressing: tombstones are silently dropped when the model needs downloading

The rebuild runs before the download block:

  // rebuild info.recipe_options with tombstones removed                                                                                                                                                                                                                                                                     
  info.recipe_options = RecipeOptions(info.recipe, per_model_options);                                                                                                                                                                                                                                                       
                                                                                                                                                                                                                                                                                                                             
  if (!model_manager_->is_model_downloaded(model_name) && !is_model_collection_recipe(info.recipe)) {                                                                                                                                                                                                                        
      ...                                                                                                                                                                                                                                                                                                                    
      info = model_manager_->get_model_info(model_name);   // <-- overwrites the tombstoned recipe_options                                                                                                                                                                                                                   
  }                                                                                                                                                                                                                                                                                                                          

If the model isn't already downloaded, info is re-fetched from get_model_info(), which rebuilds recipe_options from the full saved layer (including tombstoned keys). The transient null clear then has no effect — the model loads with the saved value. This is reachable because options can be saved without loading
(POST /models/{id}/options) and download state is independent of the saved options. It affects both the save_options and non-save_options paths.

Fix suggestion: move the tombstone-rebuild to after the download block (after both info reassignments), so it is applied to the final info used by router_->load_model.

Minor note: the tombstone-rebuild block and the save-write preservation block both re-read get_saved_model_options and could be factored into a small helper, but this is cosmetic.

[fl0rianr]

Good catch. Moved the transient tombstone overlay after the download/re-fetch so it is applied to the final ModelInfo passed to the router. Kept this as a minimal ordering fix without factoring the surrounding logic.