intword: values between 10^36 and a googol format as huge decillion counts, and rounding to the next unit does not carry

從報告中提到的 number.py 變更開始，並使用 Python 3.11 重現邊界範例。向 maintainers 確認該缺口應回退到 str(value)，還是在 googol 邊界繼續處理，然後為選定的行為新增回歸測試。當接近 10^100 的值不再產生巨大的 decillion 計數，且捨入遵循議定的規則時，即表示完成。

What happened

intword handles values up to 1000 decillion (10^36) and exactly a googol
(10^100) well, but everything in between formats as an enormous decillion
count:

>>> import humanize
>>> humanize.intword(10**100)
'1.0 googol'
>>> humanize.intword(10**100 - 10**93)
'9999999000000000610151082626636586748864734715312950532670947852288.0 decillion'
>>> humanize.intword(10**50)
'100000000000000000.0 decillion'

A 67 digit unit count defeats the purpose of a human readable number, and
a value just below a googol that rounds up to it never carries to
"1.0 googol".

Verified on humanize 4.16.0 (published wheel) and on current main
(c3a124c), Python 3.11.

Why this looks like a bug rather than a choice

This is the same class of boundary behavior that #328 recently fixed for
metric() (carry to the next SI prefix when rounding reaches 1000). The
powers list jumps from decillion (10^33) to googol (10^100), and the
selection loop picks decillion for anything under a googol, no matter how
large the resulting count is. Values above the largest power already fall
back to str(value), which suggests the intended behavior for
"no reasonable unit" is the plain number, not a unit with a 67 digit
multiplier.

Expected

One of the following, whichever matches the project's intent:

1. Values whose formatted count reaches 1000 of a unit with no adjacent
   next power fall back to str(value) (consistent with the existing
   above-maximum behavior), and values that round up to exactly the next
   power carry to it ("1.0 googol"), consistent with #328's carry rule.
2. Or simply fall back to str(value) for the whole gap between 10^36
   and 10^100.

I am happy to open a PR with regression tests for the boundary cases once
maintainers confirm which behavior is intended.

How this was found

An automated review gate proposed edge case tests for the recent number.py
changes (the #328 carry fix among them) and executed them against the
code; the "carry just below the next unit" test failed with the output
above. The failing test and its observed output are reproducible by hand
as shown.