https://hackerone.com/reports/3437743

The method incorrectly ignored its first argument and treated the second argument as offset and the third as count.
This change makes the first argument be treated as offset and the second as count.

Also fix incorrect block parameter in comments.

- Buffer's size did not account for offset when mapping the file, leading to possible crashes.
- Size and offset were not checked properly, leading to many situations raising EINVAL errors with generic messages.
- Documentation was wrong.

The `source` field in IO::Buffer can have a String or an IO::Buffer
object, if not nil.

- When the `source` is a String object. The `base` field points to the
  memory location of the String content, which can be embedded in
  RSTRING, and in that case, GC compaction can move the memory region
  along with the String object.

  Thus, IO::Buffer needs to pin the `source` object to prevent `base`
  pointer from becoming invalid.

- When the `source` is an IO::Buffer, then `base` is a pointer to a
  malloced or mmapped memory region, managed by the source IO::Buffer.
  In this case, we don't need to pin the source IO::Buffer object,
  since the referred memory region won't get moved by GC.

Closes: [Bug #21210]

The current implementation of `IO::Buffer#copy` and `#set_string` has
an undefined behavior when the source and destination memory overlaps,
due to the underlying use of the `memcpy` C function.

This patch guarantees the methods to be safe even when copying between
overlapping buffers by replacing `memcpy` with `memmove`,

Fixes: [Bug #20745]