PERF: creating string Series/Arrays from sequence with many strings

[topper-123]

Improves performance of pandas._libs.lib.ensure_string_array is cases with many string elements.

Examples:

>>> x = np.array([str(u) for u in range(1_000_000)], dtype=object)
>>> %timeit pd.Series(x, dtype=str)
344 ms ± 59.7 ms per loop  # v1.1.0
157 ms ± 7.04 ms per loop  # v1.1.1 and master
22.6 ms ± 191 µs per loop  # this PR
>>> %timeit pd.Series(x, dtype="string")
357 ms ± 40.2 ms per loop  # v1.1.0
148 ms ± 713 µs per loop  # v1.1.1 and master
26.3 ms ± 291 µs per loop  # this PR

#35519 is the cause of the improvement from 1.1.0 to 1.1.1.

Together with #35519 this PR means that the overhead of working with strings in pandas has gotten considerably smaller for cases when we many times instantiate new Series/PandasArrays with dtype str/StringDtype, i.e. probably quite often.

[jbrockmendel]

nice. Does this have/merit an asv?

[jreback]

nice, ping when ready (if we need an asv) or if we have sufficient.

[topper-123]

There are ASVs for both strand "string"in asv_bench/benchmarks/strings.py.

[jreback]

thanks @topper-123 very nice!