[cpruijsen]

Title: FIX: Pass a mask image to N4 so --rescale-intensities takes effect

Connects a brain mask to `mask_image` at every N4 node that sets `rescale_intensities=True`, so the rescaling takes effect. Fixes #1054.

`N4BiasFieldCorrection` applies `--rescale-intensities` (`-r`) only when a mask image (`-x`) is given. Without one, ANTs sets `doRescale = false` and ignores the flag. `weight_image` (`-w`) does not count as a mask for this check.

Each node now uses a mask the workflow already computes:

| Workflow | N4 node | `mask_image` source |
|---|---|---|
| `init_brain_extraction_wf` (no Atropos refinement) | `inu_n4_final` | `thr_brainmask`, the warped and thresholded template brain mask (also `out_mask`) |
| `init_atropos_wf` | `inu_n4_final` | `msk_conform`, the refined brain mask (also `out_mask`) |
| `init_n4_only_wf` | `inu_n4` | `binarize`, the binarized first input (also `out_mask`) |
| `init_enhance_and_skullstrip_bold_wf` | `n4_correct` | the tentative mask that already feeds `weight_image` (both `pre_mask` branches) |

This matches the N4 step in ANTs' `antsAtroposN4.sh`, which passes the brain mask with `-x` and the weights with `-w`.

In `afni_wf` (`niworkflows/anat/skullstrip.py`), N4 runs before skull-stripping, so no mask exists. The flag did nothing there, so I removed it.

`-x` also limits the region where N4 fits the bias field, so the `bias_corrected` outputs of these workflows change numerically. `init_n4_only_wf` changes most, because N4 used the whole image before. The no-change alternative is to drop `rescale_intensities=True` from all five nodes, which abandons the intent of 406b4e5. I chose the mask because rescaling is what that setting was for. If you prefer the no-change option, or a different mask at any node, I can update the PR.

New construction-time tests in `niworkflows/anat/tests/test_ants.py`, `niworkflows/anat/tests/test_skullstrip.py` and `niworkflows/func/tests/test_util.py` check the flag and the `mask_image` source for each node. All 7 fail on `master` and pass with this change. I did not run ANTs on real data; the issue shows the runtime effect (output range 0 to 9999 with `-x`, 0 to 14617 without).