Various IO::Buffer improvements.

[ioquatix]

https://bugs.ruby-lang.org/issues/18417

• IO::Buffer.map('README.md') (just a string, I experimented "whether it is smart enough") core dumps.
• I suspect we might want to set buffer to IMMUTABLE if it is created from the frozen string. Currently, this works: (demo of broken behaviour changing frozen string contents).
• The buffer raises RuntimeError as the only type of the error. I believe that in some places, it should be other standard errors (for example, "Offset + length out of bounds" could be at least ArgumentError?..), and in other occurrences, special error classes: for example, "Buffer is locked" would probably be helpful to have as some IO::Buffer::LockedError.
• It seems to_str always returns a string in US-ASCII encoding, I am not sure it is the right behavior. For a low-level buffer, I'd rather expect ASCII-8BIT (what file read in binary mode produces in Ruby).
• (Not sure if a bug or just not very useful behavior?) #resize applied to a buffer mapped from file seems to "break" the link between the buffer and file; no further copy or set would be reflected on file.
• It seems that #clear behaves weirdly when offset is provided but length is not.

[ioquatix]

(Not sure if a bug or just not very useful behavior?) #resize applied to a buffer mapped from file seems to "break" the link between the buffer and file; no further copy or set would be reflected on file.

I've reworked the EXTERNAL flag to handle this case better. Now the buffer can be EXTERNAL|MAPPED which means we can't resize it. Resizing this is confusing because it breaks the external mapping (visibility of changes). I think this is a more logical place.

If someone ACTUALLY wants to resize the file, they need to use ftruncate and then REMAP the file. We can try this operation later, specifically handling this case, but it will invalidate all references/slices to the buffer.