Drop log handling in sliced epochs not intuitive

Describe the bug

Using drop_log_stats() or drop_log is unintuitive when working with sliced epochs.

Steps to reproduce

MWE

import mne
import numpy as np

np.random.seed(10)
n_epochs, n_channels, n_timepoints = 10, 4, 21
ch_names = [f"ch_{i}" for i in range(n_channels)]
fake_data = 1e-6*np.random.random((n_epochs, n_channels, n_timepoints))

# Have the first three epochs exceed minmax in one channel
fake_data[0:3, 1, 0] = 200e-6

info = mne.create_info(ch_names=ch_names, sfreq=40)
epo = mne.EpochsArray(fake_data, info)
# This should drop the first three epochs
epo.drop_bad(reject=dict(misc=100e-6))

# Prints 30% as expected
print(f"I expect 30% and am: {epo.drop_log_stats()}%")

# select only the last epoch
subepos = epo[-1:]
# Prints 75%
print(f"I expect 0% or maybe 30% but am: {subepos.drop_log_stats()}")

# Drop log contains IGNORED only for events that were not dropped initially
print(subepos.drop_log)
# Output is:
# (('ch_1',), ('ch_1',), ('ch_1',), ('IGNORED',), ('IGNORED',), ('IGNORED',), ('IGNORED',), ('IGNORED',), ('IGNORED',), ())

Expected results

I would either expect the drop log stats to know the initial ratio of drops, i.e. 30%, or after slicing report only the dropped epochs in the slice, i.e. 0%. However, calculating the latter is hard for non-trivial drop/slice combinations.

Actual results

MNE remembers how many epochs were dropped initially, then ignores the epochs marked as IGNORED and calculates the drop log stats to be 75%, as 3 were dropped and the slice contains 1 non-dropped epoch.

I suspect that when doing slices of epochs, the __getitem__ implementation only inserts ('IGNORED') in the drop log when there is no entry yet.

Additional information

Platform:      Linux-5.11.0-34-generic-x86_64-with-glibc2.29
Python:        3.8.10 (default, Jun  2 2021, 10:49:15)  [GCC 9.4.0]
Executable:    /home/jan/dev/python/spot/spot_venv/bin/python
CPU:           x86_64: 8 cores
Memory:        Unavailable (requires "psutil" package)
mne:           0.24.dev0
numpy:         1.21.2 {blas=openblas, lapack=openblas}
scipy:         1.7.1
matplotlib:    3.4.3 {backend=module://backend_interagg}
sklearn:       0.24.2
numba:         Not found
nibabel:       3.2.1
nilearn:       Not found
dipy:          Not found
cupy:          Not found
pandas:        1.3.2
mayavi:        Not found
pyvista:       Not found
vtk:           Not found

Start by reproducing the MWE, then trace Epochs.getitem and the drop_log_stats() and drop_log behavior for sliced epochs. The expected semantics need to be settled between initial-drop statistics and slice-only statistics; done means the chosen behavior is documented and covered by regression tests.