PERF: improve to_timedelta perf for string-like

[jreback]

see #6754
related #3009, #8190
from SO

convering string like timedeltas, e.g.

'1 days', '3 days, 00:00:03', '3 days 2 hours'

is slow because it requires a regex match for each element. Can put cythonized (like format parsing for datetimes), or since only very limited formats supported can simply test them (in cython) and convert.