Fix multiple bugs in IO::Buffer.map and update its documentation

[trinistr]

Main issues that are fixed:

• Documentation was very outdated to the point that it was straight-up wrong.
• Buffer's size didn't account for offset, leading to crashes on something like

  buffer = IO::Buffer.map(File.open("data.txt"), nil, IO::Buffer::PAGE_SIZE, IO::Buffer::READONLY)
  data = buffer.get_string

Other issues were about not checking arguments enough, leading to a proliferation of Errno::EINVAL with generic messages:

• Empty file and/or size of 0 was allowed.
  • This is a breaking change, as this worked on Windows. This seems entirely useless, however, as nothing can be done with such a buffer, as far as I'm aware.
• Size and offset could be larger than the file.
• Offset was not checked for negative values.

I've also added a bunch of tests. Note, however, that there are no tests for successfully using offset, as it's impossible to reliably use (see https://bugs.ruby-lang.org/issues/21700).

[trinistr]

The test failure doesn't seem connected with my changes in any way and didn't happen before fixing style.

1. Failure:
   TestTimeExtension#test_rfc2822_nonlinear [/Users/runner/work/ruby/ruby/src/test/test_time.rb:67]:
   [ArgumentError] exception expected, not #<Timeout::Error: [5000]: in 0.19219318140835537s (tmin: 1.1790999999983232e-05, tmax: 5.9792000000002954e-05, tbase: 0.0038438636281671076)>.

[ioquatix]

Suggested change

	if (UNLIKELY(file_size <= 0)) {
	if (UNLIKELY(file_size < 0)) {

In wonder if we should allow mapping zero-sized files? Not sure what the point would be, but is it technically invalid?

[ioquatix]

I think it only matters if the mapped size is going to be too big (bigger than SIZE_MAX).

Also, I believe RB_NUM2SIZE will already raise if out of range, e.g.:

unsigned long
rb_big2ulong(VALUE x)
{
    unsigned long num = big2ulong(x, "unsigned long");

    if (BIGNUM_POSITIVE_P(x)) {
        return num;
    }
    else {
        if (num <= 1+(unsigned long)(-(LONG_MIN+1)))
            return -(long)(num-1)-1;
    }
    rb_raise(rb_eRangeError, "bignum out of range of unsigned long");
}

[ioquatix]

I mostly agree with this change. I may make some follow up refinements once this is merged.