Fix IrregularlySampledSignal.time_slice returning the whole signal for an empty window

[Yi-111-a]

Fixes #1888.

The bug

IrregularlySampledSignal.time_slice computes a correct boolean mask, but then
recovers the integer slice bounds by walking it:

id_start = None
id_stop = None
for i in indices:
    if id_start is None:
        if i:
            id_start = count
    else:
        if not i:
            id_stop = count
            break
    count += 1

new_st = deepcopy(self[id_start:id_stop])

None carries two meanings here: "no match found yet" inside the loop, and
"unbounded" in the slice afterwards. When no sample falls inside the window both
stay None, self[None:None] evaluates to self[:], and the entire signal is
returned.

Segment.time_slice appends the result unconditionally, so a segment sliced to a
window in which one irregularly sampled channel has no samples comes back holding
that channel's whole recording — timestamped outside the window that was asked
for — while its events and epochs for the same window are correctly dropped.

Reproducer against current master

import numpy as np
import quantities as pq
from neo.core import IrregularlySampledSignal

times = np.array([1.0, 2.0, 30.0, 40.0]) * pq.s
sig = IrregularlySampledSignal(times, np.arange(4.0).reshape(-1, 1) * pq.mV)
print(sig.time_slice(5 * pq.s, 20 * pq.s).times)   # [ 1.  2. 30. 40.] s

The fix

Return an empty signal when nothing matched:

if id_start is None:
    return deepcopy(self[0:0])

I chose "empty" over raising ValueError deliberately — as the issue says, the
sibling classes split on this and it is a real design question. I went with empty
because:

• Event.time_slice and Epoch.time_slice return empty for the same situation;
• an empty IrregularlySampledSignal is already a state this class supports. The
  constructor accepts one, and test_time_slice_empty already asserts that
  slicing one returns an empty one. The current code satisfies that test only by
  accident, because its input is empty to begin with;
• AnalogSignal.time_slice raises, and Segment.time_slice propagates that, so
  raising here would widen what can fail inside a segment slice. Returning empty
  changes behaviour only in the case that is currently wrong.

If you would rather raise here, I am happy to rework it — the mask is already
correct, so it is a one-line change either way.

I left Segment.time_slice alone. It still appends the now-empty signal
unconditionally; whether a channel with no samples should be dropped like an empty
Event seems worth deciding separately, and it is not needed to fix the reported
behaviour.

Verification

• pytest neo/test/coretest/test_irregularysampledsignal.py — 70 passed, 1 skipped
• pytest neo/test/coretest/ — 621 passed, 11 skipped
• The new test_time_slice_window_without_samples fails on master with
  AssertionError: [ 1. 2. 30. 40.] s != [] s and passes with the fix.
• Windows that do contain samples are unaffected: time_slice(None, None),
  time_slice(3, None), time_slice(None, 3), and the existing
  test_time_slice_out_of_boundries all still pass.