h/t Ron Minnich

See https://groups.google.com/a/groups.riscv.org/forum/#!topic/isa-dev/JV-Hj3W5Kw8

Currently the DESTDIR variable is not used correctly which leads to
bogus RUNPATH entries.

https://www.gnu.org/prep/standards/html_node/DESTDIR.html

The flag is 32 bits, and if we read 64/128 bits then we get fragments of
S1 too and can accidentally send an error.  Fixes #84.

Fixes 32-bit build.

It seems that g++ 5.4 doesn't realize that it is building a precompiled 
header unless you pass it -x c++-header.

This commit also factors out the common AMO code into mmu_t.

Temporarily turn them on in mstatus if necessary.

It was erroneously complaining that gdb sent too much data even when it
wasn't.

Newer gdbs send larger memory write packets when downloading.
Also improve error reporting when gdb sends packets that don't fit in
the buffer.

Allow the CSR writer to make the choice.

@timsifive @colinschmidt this fixes the failing rv64mi-p-breakpoint test.

Instead, just give gdb what it asks for.
Also when gdb does a register write, let the user know that it's likely
misconfigured and tell them how to fix it.

This is probably as well as issue #72 can be fixed in spike.

M-mode writes to tdata1 with dmode set are ignored instead of raising an
exception.
Add the same behavior for tdata2.

They live in riscv-tests/debug now, since they also test gdb, and can be
used to test other targets besides spike.

Implement address and data triggers.

Conflicts:
	riscv/encoding.h
	riscv/processor.cc

Add timing bit (but it doesn't do anything).
Implement dmode bit.