offset_times with an array n truncates offsets finer than the times' resolution #304

For non-Tick offsets, _offset_times_array_n casts the result back to the input resolution. An offset finer than that resolution is silently dropped with an array n, while a scalar n keeps it:

import numpy as np
import pandas as pd
from utilsforecast.processing import offset_times

s = pd.Series(pd.to_datetime(["2020-01-01", "2020-01-31"])).astype("datetime64[s]")
off = pd.DateOffset(milliseconds=500)
offset_times(s, off, 1)                 # 00:00:00.5, datetime64[ms]
offset_times(s, off, np.array([1, 1]))  # 00:00:00,   datetime64[s]

This path also emits a pandas FutureWarning ("The behavior of array concatenation with empty entries is deprecated") from dt_index[:0].append(runs) when the runs have a finer unit than the input.

Proposed fix: drop the final cast in _offset_times_array_n so an array n keeps the computed resolution, as a scalar n and the Tick path already do. This changes the output dtype for these inputs.

time_ranges has a similar truncation (it casts to starts.dtype), e.g. time_ranges(<datetime64[s] series>, "500ms", 2) repeats each start.

Found while reviewing #303.