fetch1: Change the way predictions from the BTC are sent downstream

fetch1: Change the way predictions from the BTC are sent downstream

Instead of sending down the predicted taken/not-taken bits with the
target of the branch, we now send them down with the branch itself.
Previously icache adjusted for this by sending the prediction bits to
decode1 without a 1-clock delay while everything else had a 1-clock
delay.  Now icache keeps the prediction bits with the rest of the
attributes for the request.

Also fix a buglet in fetch1 where the first address sent out after
reset didn't have .req set.  Currently this doesn't cause a problem
because icache doesn't really look at .req.

Signed-off-by: Paul Mackerras <paulus@ozlabs.org>