Fix Fiddle.last_error on FFI backend and improve test to work for all

[eregon]

No description provided.

[eregon]

Mmh, JRuby on Windows seems to have a bug here, cc @headius:
https://github.com/ruby/fiddle/actions/runs/13081426763/job/36505655327?pr=173#step:7:31

assert_equal RbConfig::LIMITS["LONG_MAX"], strtol.call((2**128).to_s, nil, 10)
<9223372036854775807> expected but was
<2147483647>

LONG_MAX on Windows is 2147483647 (LLP64), but RbConfig::LIMITS["LONG_MAX"] on JRuby on Windows reports 9223372036854775807 (incorrect).
(I made sure the value before expected is the LHS, so it's not the other way around)

[kou]

jruby/jruby#8357 may be related for the JRuby behavior.

[eregon]

Worked around the JRuby on Windows failure by not asserting the result of strol(), it's not really important, we are testing errno.

[eregon]

There is another error for JRuby on Windows, I'll skip that test for it then:
https://github.com/ruby/fiddle/actions/runs/13082143360/job/36507738494?pr=173

[eregon]

CI is green now and this PR is ready

[kou]

Thanks.