naturaltime and precisedelta both routed their float input through
_date_and_delta, which fed it straight to round() (or skipped round
in precise mode) and then dt.timedelta(seconds=...). For a non-finite
float like inf or -inf, round(float("inf")) raises OverflowError
and dt.timedelta(seconds=float("inf")) raises OverflowError too.
NaN happened to work because round(float("nan")) raises ValueError
which the except clause caught, but inf/-inf slipped past.

The naturaldelta path caught inf/-inf via a different probe (it
calls int(value) first, which raises OverflowError for inf) and
that case is fixed in the still-open PR python-humanize#342 -- this commit leaves
that probe alone and adds the equivalent isfinite guard inside
_date_and_delta, which is what both naturaltime and precisedelta
go through. Adding OverflowError to the except tuple covers the
last remaining case where a finite but oversized number
(> ~3e9 days in seconds) would otherwise crash.

Three new parametrized test functions in tests/test_time.py cover
naturaltime, precisedelta, and the internal _date_and_delta
contract for inf/-inf/nan. The inf and -inf cases fail on the
pre-fix code with OverflowError; the nan case already worked and
is included as a regression guard. Full suite: 724 passed,
69 skipped.