Annotation timing info loss due to datetime.timedelta precision limit

Description of the problem

Sub-microsecond timing data is lost when pushed through a datetime.timedelta object step, for example in mne.annotations.crop. As raw.set_annotations calls crop internally, this can lead to misalignment of the created annotations and raw.times, with the annotations covering slightly different frames than expected.

This will only affect cases where 1/sfreq has sub-microsecond precision, but 512Hz is common in my experience, and 1/512 = 0.001953125s.

Steps to reproduce

from datetime import datetime, timezone

import numpy as np
from mne import Annotations, create_info
from mne.io import RawArray

info = create_info(5, 512, "eeg")
data = np.random.randn(5, 10000)
raw = RawArray(data, info)
print(raw.times[1])
# 0.001953125
annotation = Annotations(onset=raw.times[1], duration=1, description="test")
print(annotation[0])
# OrderedDict([('onset', 0.001953125), ('duration', 1.0), ('description', 'test'), ('orig_time', None)])
raw.set_annotations(annotation)
print(raw.annotations[0])
# OrderedDict([('onset', 0.001953), ('duration', 1.0), ('description', 'test'), ('orig_time', None)])
print(raw.time_as_index(raw.annotations[0]['onset']))
# array([0])

Link to data

No response

Expected results

The annotation details shouldn't be lost when attached to the raw object. The correct original range of frames should be as precisely obtainable from the timing data as before attached to the raw object.

annotation[0]['onset'] and raw.annotations[0]['onset'] should be equal at 0.001953125.
raw.time_as_index(raw.annotations[0]['onset']) should return array([1]), corresponding to how the annotation was created.

Actual results

annotation[0]['onset'] is 0.001953125 while raw.annotations[0]['onset'] is 0.001953.
raw.time_as_index(raw.annotations[0]['onset']) returns array([0])

Additional information

Platform Linux-3.10.0-1160.102.1.el7.x86_64-x86_64-with-glibc2.17
Python 3.11.5 (main, Sep 5 2023, 13:23:33) [GCC 4.8.5 20150623 (Red Hat 4.8.5-44)]
Executable /isilon/LFMI/VMdrive/Jonathan/PriorWeightingMooney/Visual/ECoG/Code/.venv/bin/python
CPU x86_64 (20 cores)
Memory 96.2 GB

Core
├☑ mne 1.5.1
├☑ numpy 1.26.2 (OpenBLAS 0.3.23.dev with 20 threads)
├☑ scipy 1.11.4
├☑ matplotlib 3.8.2 (backend=module://matplotlib_inline.backend_inline)
├☑ pooch 1.8.0
└☑ jinja2 3.1.2

Numerical (optional)
├☑ sklearn 1.3.2
├☑ nibabel 5.1.0
├☑ nilearn 0.10.2
├☑ pandas 2.1.3
└☐ unavailable numba, dipy, openmeeg, cupy

Visualization (optional)
├☑ pyvista 0.42.3 (OpenGL unavailable)
├☑ pyvistaqt 0.11.0
├☑ vtk 9.3.0
├☑ ipympl 0.9.3
├☑ ipywidgets 7.8.1
└☐ unavailable qtpy, pyqtgraph, mne-qt-browser, trame_client, trame_server, trame_vtk, trame_vuetify

Ecosystem (optional)
└☐ unavailable mne-bids, mne-nirs, mne-features, mne-connectivity, mne-icalabel, mne-bids-pipeline

Start at mne/annotations.py around Annotations.crop and trace how raw.set_annotations passes timing values through it. Run the reproduction in the issue, then verify that attached annotation onset retains 0.001953125 and raw.time_as_index returns array([1]) without losing sub-microsecond precision.