include_images=false still enables image generation for non-placeholder export modes

## Problem

`include_images=false` does not currently suppress image generation when `image_export_mode` is `embedded` or `referenced`.

In the conversion manager, `image_export_mode != placeholder` enables page image generation, and `referenced` additionally enables picture image generation, without checking `include_images`.

That makes the two options conflict semantically: callers can explicitly ask not to include images, but still get generated/persisted image assets because the export mode is referenced.

## Expected

`include_images=false` should be the controlling option for generated image assets. If it is false, conversion should not enable page or picture image generation regardless of `image_export_mode`.

## Evidence

In a downstream docling-jobkit/docling-serve workload on a 157-page PDF:

- `include_images=false` with `image_export_mode=referenced` still emitted 486 image assets, about 42.8 MB of assets, and about 130 MB JSON.
- `include_images=false` with `image_export_mode=placeholder` emitted 0 assets and about 14.2 MB JSON.

The table/chunk text outputs were stable across the two runs, so the extra assets came from image-generation/export semantics rather than document content.

## Proposed fix

Gate page and picture image generation on `include_images` before applying `image_export_mode`.

Start in the conversion manager where image_export_mode enables page and picture image generation. Trace how include_images and image_export_mode are applied, then verify that include_images=false prevents generated image assets for embedded, referenced, and placeholder modes while preserving the expected conversion outputs.

Written by the indexing model from the issue text.