[Perf] Refactor tensor disposal logic to reduce memory usage (#966)

### What this PR does / why we need it?
1. In previous PRs https://github.com/vllm-project/vllm-ascend/pull/580
https://github.com/vllm-project/vllm-ascend/pull/784, I saved GPU memory
by promptly deleting unnecessary tensors. For tensors passed from
upper-layer functions, I used a list container to transfer the parameter
and then popped the tensor from the list within the inner function to
achieve deletion. Recently, I discovered a better implementation in
sglang—the `dispose_tensor` function and I recommend adopting this
approach.
2. Dispose `hidden_states` and `residual` from the previous layer once
they're no longer used.
3. Avoid to generate `self.inputs_embeds` in `ModelRunnerV1` in
non-multimodal scenarios.

With the aforementioned optimizations, using the DeepSeek-R1-W8A8 model
under the conditions of `TP=16` and `max-model-len=32768`, we can save
1.3GB of npu memory.

**Reference**: https://github.com/sgl-project/sglang/pull/6147

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?

---------

Signed-off-by: ApsarasX <apsarax@outlook.com>

vllm_ascend/models/deepseek_v2.py

@@ -68,6 +68,7 @@ from vllm.sequence import IntermediateTensors
 import vllm_ascend.envs as envs_ascend
 from vllm_ascend.ops.fused_moe import AscendFusedMoE
 from vllm_ascend.quantization.w8a8_dynamic import AscendW8A8DynamicLinearMethod
+from vllm_ascend.utils import dispose_tensor
 
 VLLM_ENABLE_MC2: bool = envs_ascend.VLLM_ENABLE_MC2
 
@@ -518,8 +519,14 @@ class CustomDeepseekV2DecoderLayer(DeepseekV2DecoderLayer):
             residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
         else:
+            previous_hidden_states, previous_residual = hidden_states, residual
             hidden_states, residual = self.input_layernorm(
                 hidden_states, residual)
+            # Dispose hidden_states and residual from the previous layer
+            # to save npu memory because they're no longer used.
+            dispose_tensor(previous_hidden_states)
+            dispose_tensor(previous_residual)
+
         hidden_states = self.self_attn(
             positions=positions,
             hidden_states=hidden_states,