Read less data from the input connection.

Currently, we read 256 (max word size) * 5 (max N-gram size + 1) characters
from the input connection when building our context. This is overkill. We
don't need more than 80 characters, regardless of which decoder we use.

Bug 19987461.

Change-Id: Ie3a321cf2482adbacd8006d9d86e6601097c15ed