qwen35: fix do_spec_decode argmax OOB on prefix-cache partial restore

`n_last_chunk = committed % PREFILL_UBATCH` only equals the last prefill
chunk's actual size when prefill started at kv_offset=0. With prefix-cache
partial restore, `restore_and_generate` runs delta-prefill from kv_offset>0,
so the last chunk's `n_tokens` is `prompt_len - kv_offset`, not the modulo
of `committed` over PREFILL_UBATCH. The read offset was then larger than
sg_.argmax_tokens->ne[0], firing the "tensor read out of bounds" assert
on the first DFlash spec-decode request against any prompt the cache had
already seen.

Read the actual last-chunk size from sg_.argmax_tokens->ne[0], which the
graph builder sized to match the bound chunk. No-op when kv_offset==0
(`committed % UBATCH == ne[0]`).

Author: dusterbloom

dflash/src/qwen35/qwen35_backend.cpp

@@ -832,11 +832,17 @@ bool Qwen35Backend::do_spec_decode(int committed, int n_gen,
     // holds argmax for ALL positions in that chunk. We need the LAST position.
     int32_t last_tok;
     {
-        const int PREFILL_UBATCH = 512;
-        int n_last_chunk = committed % PREFILL_UBATCH;
-        if (n_last_chunk == 0) n_last_chunk = PREFILL_UBATCH;
+        // The last prefill chunk's size is whatever the still-bound argmax
+        // tensor says. Deriving it from `committed % PREFILL_UBATCH` was wrong
+        // when prefix-cache partial restore made the chunk shorter than the
+        // committed position would suggest (delta-prefill from kv_offset != 0).
+        const int64_t n_last_chunk = sg_.argmax_tokens ? sg_.argmax_tokens->ne[0] : 0;
+        if (n_last_chunk <= 0) {
+            std::fprintf(stderr, "do_spec_decode: argmax_tokens missing or empty\n");
+            return false;
+        }
         ggml_backend_tensor_get(sg_.argmax_tokens, &last_tok,
-                                sizeof(int32_t) * (n_last_chunk - 1),
+                                sizeof(int32_t) * (size_t)(n_last_chunk - 1),
                                 sizeof(int32_t));
     }