Preserve caller-owned memory-mapped Raw preload files

[bruAristimunha]

Reference issue (if any)

None.

What does this implement/fix?

This fixes the ownership semantics of caller-named memory-mapped preload files.

MNE already supports storing decoded data in a memory-mapped file through preload=<path> and raw.load_data(memmap=<path>). Because that path is supplied by the caller, the resulting file should remain
under caller ownership. Previously, BaseRaw.__del__ could unlink the backing file when the Raw object was garbage-collected.

This PR removes that automatic unlinking and clarifies the existing contract:

• A caller-supplied path is freshly created or overwritten; it is not reused as a cache.
• The file remains available after Raw.close() and after the Raw object is destroyed.
• The caller is responsible for removing the file when it is no longer needed.
• Caller-supplied memory-mapped arrays passed to RawArray follow the same ownership rule.

This keeps the existing preload concepts unified: passing a path means “decode into this caller-owned memory-mapped file.” It does not introduce source-file memory mapping, persistent file handles, cache
reuse, or a new preload mode.

Additional information

Tests cover:

• construction with preload=<path>;
• load_data(memmap=<path>);
• overwriting and truncating an existing destination;
• survival across close() and garbage collection;
• reopening and replacing the destination;
• concatenation, copying, and appending;
• direct RawArray memory maps;
• signal-bearing EDF and FIF inputs.

Validation performed locally:

• focused ownership tests: 7 passed, 4 testing-dataset skips;
• adjacent Raw, EDF, and FIF tests: 174 passed, 81 testing-dataset skips;
• docstring parameter tests: 19 passed;
• pre-commit hooks passed.

[larsoner]

I'm going to kill CIs on this and other PRs that depend on previous ones... otherwise no other PRs will have runners to test.

[bruAristimunha]

Makes sense — thanks. Holding all pushes to the stacked branches until #14212 merges so runners aren't consumed by re-runs; downstream drafts (#14214–#14216) likewise won't be pushed until their parents land.

[larsoner]

I think this should already work in theory:

>>> import mne
>>> raw = mne.io.read_raw_edf("/Users/larsoner/python/MNE-testing-data/EDF/SC4001EC-Hypnogram.edf", preload="test.tmp")
<python-input-1>:1: RuntimeWarning: Header information is incorrect for record length. Default record length set to 1.
It is possible that this file only contains annotations and no signals. In that case, please use mne.read_annotations() to load these annotations.
<python-input-1>:1: RuntimeWarning: Omitted 153 annotation(s) that were outside data range.
<python-input-1>:1: RuntimeWarning: Limited 1 annotation(s) that were expanding outside the data range.

But it maybe "just" needs to be fixed? Because this is not the right result...

>>> raw._data
memmap([], shape=(0, 2054), dtype=float64)

It should look more like this

>>> raw = mne.io.read_raw_fif(mne.datasets.sample.data_path() / "MEG" / "sample" / "sample_audvis_raw.fif", preload="test.tmp2")
>>> raw._data
memmap([[ 9.64355481e-12,  0.00000000e+00,  0.00000000e+00, ...,
         -1.92871096e-12,  2.89306644e-12,  3.85742192e-12],
        [-4.82177740e-12, -2.89306644e-12, -9.64355481e-13, ...,
         -9.64355481e-13, -9.64355481e-13, -1.92871096e-12],
        [ 1.01074222e-13,  6.31713890e-14,  7.58056668e-14, ...,
         -4.80102556e-13, -6.06445334e-13, -5.93811056e-13],
        ...,
        [ 3.88542173e-05,  4.07510373e-05,  4.09957883e-05, ...,
          6.72453304e-05,  6.68782039e-05,  6.91421504e-05],
        [ 6.58391126e-05,  6.80025648e-05,  6.81779798e-05, ...,
          8.51932390e-05,  8.58948991e-05,  8.89938982e-05],
        [ 2.85661012e-04,  2.83699953e-04,  2.80431520e-04, ...,
          2.64089357e-04,  2.62781984e-04,  2.57552492e-04]],
       shape=(376, 166800))

[bruAristimunha]

hey @larsoner, can you please check again?

[larsoner]

Okay for me!