Implement AsyncWrite::poll_write_vectored for BufWriter, making use
of the buffer to coalesce data from the passed slices. Make exceptions
for cases when writing directly to the underlying object is more
efficient, which differ depending on whether the underlying writer
itself supports vectored output.

Change the implementation of AsyncWrite::is_write_vectored to return
true.

Forward poll_write_vectored/is_write_vectored to the underlying object
for BufReader.

Do the same in BufStream, which has the effect of is_write_vectored
returning true because BufWriter's implementation now does.

Simplify branching and do what poll_write does: do not buffer
slices partially, optimizing for the most likely case where slices
are much smaller than the buffer, while retaining special treatment
for oversized slices.

In the implementation of AsyncWrite::poll_write_vectored
for BufWriter, the total length of the data can technically overflow
usize even with safely obtained buffer slices, since slices may
overlap.

Improve readability by updating the IOSlice vector slice as
constituent slices are gone over by the iteration, rather than
using an iterator.

Partially incorporated code suggested by @Darksonn.

Co-authored-by: Alice Ryhl <alice@ryhl.io>