`PaginatedPipeline._unload` prematurely unloads page backends, ignoring `keep_backend` flag

Hello,

In the `docling` library, the `PaginatedPipeline._unload` method unconditionally unloads the `_backend` for each page in `conv_res.pages`. This occurs regardless of the `keep_backend` flag, which is intended to prevent this behavior.

This premature unloading of the page backends can cause `AttributeError` exceptions if the page backends are accessed after the pipeline has completed its execution, for example, to re-render page images at a different resolution (my use case).

**Docling version**
2.41.0

**To Reproduce**

1. Create a custom pipeline that inherits from `VlmPipeline`.
2. In the main script, after `doc_converter.convert()` is called, attempt to access `page._backend` for any page in the `ConversionResult.pages` list.
3. The `_backend` attribute will be `None`, and any attempt to call methods on it will result in an `AttributeError`.

**Expected Behavior**

When `keep_backend` is set to `True` in a `PaginatedPipeline`, the page backends should not be unloaded by the `_unload` method, and should remain accessible after the conversion process is complete.

**Analysis**

The issue is in `docling/pipeline/base_pipeline.py`, in the `PaginatedPipeline._unload` method:

```python
def _unload(self, conv_res: ConversionResult) -> ConversionResult:
for page in conv_res.pages:
if page._backend is not None:
page._backend.unload()

if conv_res.input._backend:
conv_res.input._backend.unload()

return conv_res
```

This method does not check the `self.keep_backend` flag before unloading the page backends.

**Workaround**

A temporary workaround is to create a custom pipeline that overrides the `_unload` method to prevent it from unloading the page backends:

```python
class CustomVlmPipeline(VlmPipeline):
# ... (other methods)

def _unload(self, conv_res: ConversionResult):
# We override this method to prevent the backends from being unloaded.
pass # Or only unload the main document backend
```

**Proposed Solution**

The `PaginatedPipeline._unload` method should be modified to respect the `keep_backend` flag:

```python
def _unload(self, conv_res: ConversionResult) -> ConversionResult:
if not self.keep_backend:
for page in conv_res.pages:
if page._backend is not None:
page._backend.unload()

if conv_res.input._backend:
conv_res.input._backend.unload()

return conv_res
```

This change would align the behavior of the `_unload` method with the intended purpose of the `keep_backend` flag and prevent unexpected errors when working with page backends after the conversion process.