gemma4: fix prefix cache miss — add snapshot_save in generate/restore paths

The Gemma4 backend was missing the snapshot_save(req.snap_slot) call after
prefill. The prefix cache registered entries (confirm_inline_snap), but the
actual KV data was never stored. On subsequent requests, restore_and_generate
found snapshots_[slot].ctx == nullptr → 'bad slot' error → 0 output tokens.

Fix: add inline snapshot save after prefill in both generate() and
restore_and_generate(), matching the pattern used by Qwen3/Qwen35 backends.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: howard0su

dflash/src/gemma4/gemma4_backend.cpp

@@ -511,6 +511,16 @@ GenerateResult Gemma4Backend::generate(const GenerateRequest & req,
         return result;
     }
 
+    // Inline snapshot at snap_pos for prefix cache
+    if (req.snap_slot >= 0 && req.snap_pos > 0 && req.snap_pos <= committed) {
+        cache_.cur_pos = req.snap_pos;
+        if (snapshot_save(req.snap_slot)) {
+            std::fprintf(stderr, "[gemma4] inline-snap slot=%d cur_pos=%d\n",
+                         req.snap_slot, req.snap_pos);
+        }
+        cache_.cur_pos = committed;
+    }
+
     if (req.n_gen > 0) {
         // Try speculative decode if draft is available and temp==0
         const bool can_spec = dflash_target_
@@ -651,6 +661,16 @@ GenerateResult Gemma4Backend::restore_and_generate(int slot,
     }
     // else: prompt_len == snap_pos → no delta, committed stays at snap_pos
 
+    // Inline snapshot at snap_pos for prefix cache (new snap from this request)
+    if (req.snap_slot >= 0 && req.snap_pos > 0 && req.snap_pos <= committed) {
+        cache_.cur_pos = req.snap_pos;
+        if (snapshot_save(req.snap_slot)) {
+            std::fprintf(stderr, "[gemma4] inline-snap slot=%d cur_pos=%d\n",
+                         req.snap_slot, req.snap_pos);
+        }
+        cache_.cur_pos = committed;
+    }
+
     // Generate
     if (req.n_gen > 0) {
         const bool can_spec = dflash_target_