So far most of the people have been using existing ARM966E in the
place of ARM946E, because they have practically the same scan chains.

However, ARM946E has caches, which further complicates JATG handling
via scan-chain. this was preventing single-stepping for ARM946E when
SW breakpoints are used.

This patch thus introduces :
1) Correct cache handling on memory write
2) Possibility to flush whole cache and turn it off during debug, or
just to flush affected lines (faster and better)
3) Correct SW breakpoint handling and correct single-stepping
4) Corrects the bug on CP15 read and write, so CP15 values
are now correctly R/W