optimize gguf dequant, save mem, support Q2_K

use marlin for lm_head, lm_head only calc last token for prefill
extend context window to 19K for DeepSeek-V3/R1 within 24GB VRAM

ktransformers/models/modeling_deepseek_v3.py

@@ -1699,7 +1699,7 @@ class DeepseekV3ForCausalLM(DeepseekV3PreTrainedModel):
         )
 
         hidden_states = outputs[0]
-        logits = self.lm_head(hidden_states.to(self.lm_head.weight.device))
+        logits = self.lm_head(hidden_states[:,-1:,:])
         logits = logits.float()
 
         loss = None