ENH/PERF: pyarrow timestamp & duration conversion consistency/performance

[lukemanley]

• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/v2.1.0.rst file if fixing a bug or adding a new feature.

A few related changes:

1. ArrowExtensionArray.__getitem__(int) will now return a Timestamp/Timedelta for non-nano timestamp/duration types to be consistent with nanosecond types. Previously non-nano types returned python native datetime/timedelta.
2. ArrowExtensionArray.__iter__ will now yield Timestamp/Timedelta objects for non-nano types to be consistent with nanosecond types.
3. ArrowExtensionArray.to_numpy now allows for zero-copy for timestamp/duration types

Submitting as a single PR since there are a number of tests that require consistency across these methods and trying to split the nano/non-nano behavior from the performance improvements is tricky.

These were somewhat motivated by:

import pandas as pd
import pyarrow as pa

N = 1_000_000
arr = pd.array(range(N), dtype=pd.ArrowDtype(pa.timestamp("s")))

%timeit arr.astype("M8[s]")
# 5.29 s ± 162 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)          -> main
# 137 µs ± 4.88 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)   -> PR

%timeit pd.DatetimeIndex(arr)
# 6.31 s ± 560 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)          -> main
# 67.6 µs ± 3.03 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)  -> pr

[mroeschke]

Thanks @lukemanley