Prevent downgrading of hardware watchpoints if possible

The lazy flag of a value can tell us if the value contents are available.
But this could be either because it was simply never actually accessed,
or it tried to be accessed, and failed.
The latter are interesting for watchpoints, the former are not.

Currently it only uses lazy values if they are at the head of the value
chain, but this is fragile logic, and degrades some watchpoints to
software watchpoints where it's actually not necessary.

So this adds a new value flag 'm_fetch_lazy_failed' which tells if the
watchpoint expression actually tried to access the value and failed.
And this is used instead of the value chain location to tell if a lazy
value should be used as a watchpoint location.

Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=27423
Tested-By: Guinevere Larsen <guinevere@redhat.com>
Reviewed-by: Thiago Jung Bauermann <thiago.bauermann@linaro.org>
Approved-By: Tom Tromey <tom@tromey.com>