Separate grok_infnan() from grok_number().

Separate grok_infnan() from grok_number().

Remaining issues:

(1) would need tests, but there are two problems: [a] generating inf/nan
    reliably and testing for it from Perl level is hard (see items (2) and
    (3) below), and [b] the behavior of various systems with especially NaN
    differs (some platforms might throw SIGFPEs).

(2) toke.c:scan_number() will not call this code (via grok_number)
    because "NaN" or "Inf" do not look at all like floats to it.

(3) Even as we now recognize these forms, the native strtod()
    might not (problem of cross-portability of these exceptional
    forms: Win32 outputs e.g. "1.#INF", what Linux reading this should do,
    or conversely Linux outputs "Inf", what should Win32 do?)