llm: decouple prompt caching from context shift (#16639)

This PR separates prompt caching from the public shift request option for native llama-server requests.

Previously, shift controlled two different mechanisms:

context shifting / overflow behavior
per-request llama-server cache_prompt
That meant callers could not request shift: false without also disabling prompt caching.

Fixes #16635

llm/llama_server.go

@@ -1395,7 +1395,7 @@ func (s *llamaServerRunner) Completion(ctx context.Context, req CompletionReques
 	lsReq := llamaServerCompletionRequest{
 		Prompt:          prompt,
 		Stream:          true,
-		CachePrompt:     req.Shift,
+		CachePrompt:     true,
 		NPredict:        req.Options.NumPredict,
 		NKeep:           req.Options.NumKeep,
 		Temperature:     req.Options.Temperature,
@@ -1984,7 +1984,7 @@ func (s *llamaServerRunner) llamaServerChatRequest(req ChatRequest, stream bool)
 	body := map[string]any{
 		"messages":          messages,
 		"stream":            stream,
-		"cache_prompt":      req.Shift,
+		"cache_prompt":      true,
 		"n_predict":         req.Options.NumPredict,
 		"n_keep":            req.Options.NumKeep,
 		"temperature":       req.Options.Temperature,