MRG: Improve memory footprint of _TimeViewer

[GuillaumeFavelier]

This PR improves the memory footprint of _TimeViewer by decreasing the reference count and solving reference cycles.

This is the comparison between master (black curve) and this PR (blue curve) using the following script with mprof:

import os
import mne
import pyvista
from mne.datasets import sample
data_path = sample.data_path()
sample_dir = os.path.join(data_path, 'MEG', 'sample')
subjects_dir = os.path.join(data_path, 'subjects')
fname_stc = os.path.join(sample_dir, 'sample_audvis-meg')
stc = mne.read_source_estimate(fname_stc, subject='sample')
initial_time = 0.1
mne.viz.set_3d_backend('pyvista')
for j in range(5):
    for i in range(5):
        brain = stc.plot(subjects_dir=subjects_dir, initial_time=initial_time,
                         clim=dict(kind='value', pos_lims=[3, 6, 9]),
                         hemi='lh',
                         time_viewer=True, show_traces=True)
    pyvista.close_all()

There is probably room for improvement but this is a step forward.

[GuillaumeFavelier]

I don't think I can go further with this method, there is the latest result:

[larsoner]

Failure is just a timeout, thanks @GuillaumeFavelier