golang.org/cl/37258 was committed to fix issue #16146.

This patch seemed intent to allow at most one dangling byte.  But, as
implemented, many more bytes may actually slip through.  This is because
the LZW layer creates a bufio.Reader which will itself consume data
beyond the end of the LZW stream, and this isn't accounted for anywhere.

This change means to avoid the allocation of the bufio.Reader by making
blockReader implement io.ByteReader.  Further, it adds a close() method
which detects extra data in the block sequence.  To avoid any
regressions with poorly encoded GIFs which may have worked accidentally,
there are no restrictions on how many extra bytes may exist in the final
full sub-block that contained LZW data.  If the end of the LZW stream
happened to align with the end of a sub-block, at most one more
sub-block with a length of 1 byte may exist before the block terminator.

This change aims to be at least as performant as the prior
implementation.  But the primary gain is avoiding the allocation of a
bufio.Reader per frame:

name      old time/op    new time/op    delta
Decode-8     276µs ± 0%     275µs ± 2%    ~     (p=0.690 n=5+5)

name      old speed      new speed      delta
Decode-8  55.9MB/s ± 0%  56.3MB/s ± 2%    ~     (p=0.690 n=5+5)

name      old alloc/op   new alloc/op   delta
Decode-8    49.2kB ± 0%    44.8kB ± 0%  -9.10%  (p=0.008 n=5+5)

name      old allocs/op  new allocs/op  delta
Decode-8       269 ± 0%       267 ± 0%  -0.74%  (p=0.008 n=5+5)

Change-Id: Iec4f9b895561ad52266313fbc73ec82c070c3349
Reviewed-on: https://go-review.googlesource.com/68350
Run-TryBot: Emmanuel Odeke <emm.odeke@gmail.com>
Reviewed-by: Nigel Tao <nigeltao@golang.org>