LTX2.3 Text Encoder's attention proccess error and my patch

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

**Gemma**'s Text Encoder must proccess whole propmpt, not only last 1024 tokens.

### Actual Behavior

When len of whole prompt over 1024 tokens, **gemma** forgot start of prompt, because it is illuminated by 1024 attention sliding window. Other LLMs don't use attention sliding window and free out of this error.

I path this by apply window's cut only for token generation phase, not for prompt preparing:
File llama.py (changes market as #RVV):

```
class Attention(nn.Module):
#........ (skipped part) ......
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
freqs_cis: Optional[torch.Tensor] = None,
optimized_attention=None,
past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
sliding_window: Optional[int] = None,
):
batch_size, seq_length, _ = hidden_states.shape

xq = self.q_proj(hidden_states)
xk = self.k_proj(hidden_states)
xv = self.v_proj(hidden_states)

xq = xq.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
xk = xk.view(batch_size, seq_length, self.num_kv_heads, self.head_dim).transpose(1, 2)
xv = xv.view(batch_size, seq_length, self.num_kv_heads, self.head_dim).transpose(1, 2)

if self.q_norm is not None:
xq = self.q_norm(xq)
if self.k_norm is not None:
xk = self.k_norm(xk)

xq, xk = apply_rope(xq, xk, freqs_cis=freqs_cis)

present_key_value = None
if past_key_value is not None:
index = 0
num_tokens = xk.shape[2]
is_generation = num_tokens == 1 #RVV

if len(past_key_value) > 0:
past_key, past_value, index = past_key_value
if past_key.shape[2] >= (index + num_tokens):
past_key[:, :, index:index + xk.shape[2]] = xk
past_value[:, :, index:index + xv.shape[2]] = xv
xk = past_key[:, :, :index + xk.shape[2]]
xv = past_value[:, :, :index + xv.shape[2]]
present_key_value = (past_key, past_value, index + num_tokens)
else:
xk = torch.cat((past_key[:, :, :index], xk), dim=2)
xv = torch.cat((past_value[:, :, :index], xv), dim=2)
present_key_value = (xk, xv, index + num_tokens)
else:
present_key_value = (xk, xv, index + num_tokens)

if is_generation: #RVV - cut cache only in_generation
if sliding_window is not None and xk.shape[2] > sliding_window:
xk = xk[:, :, -sliding_window:]
xv = xv[:, :, -sliding_window:]
if attention_mask is not None:
attention_mask = attention_mask[..., -sliding_window:]
else:
print( "RVV PATCH - skip attentiion sliding_window:", sliding_window, ", tokens stay:", num_tokens )

xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)

output = optimized_attention(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True)
return self.o_proj(output), present_key_value

```

### Steps to Reproduce

Use long prompt with specific system instruction in begin of prompt - starting with some text length first part of instuctions will be foggoten.

### Debug Logs

```powershell
Here is not important information in terminal for this error.
```

### Other

I test my patch and it works well for long