On my Mac Mini, calling `poll()` on stdin takes around 10 us, and we
are invoking it every 20 us or so.  Reduce the frequency of polling
by 16x when not actively receiving data, thereby reducing the fraction
of time spent in `poll()` to a trivial amount.

The two-thread approach was originally motivated by making Spike look
as similar as possible to other HTIF targets.  But we can get the same
semantics without threading by running the simulator inside of the HTIF
host's idle loop instead of performing a context switch.

This was motivated by speeding up the simulator on Mac OS (it's worth
around 20% because using pthread condition variables to force strict
alternation is very slow).  But I think it also simplifies the control
flow enough to justify it on that basis, too.

Automatically test that external debug still works.

Fix supported debug version, use extension names

Addresses #1221

this is to properly handle potential 64-bit overflow in (base + size)
expression

The method can simplify proper processing of sitiations when
(base + size) overflows 64-bit interger.

NFT. We also mark `base` and `size` fields as private.

NFC. The intention is for `base` and `size` fields of mem_cfg_t
to be private members. This is the fist part of this commit.

Only update histogram when histogramming

It dates back to when this code was ifdef'd.

This is worth a 1.4x speedup on the slow path (when not histogramming).

Control mmu-dirtying via command line

Fix debug-mode regression introduced by 20e7f536

triggers: refactor: add trap_common_t

 - Debug mode should break the processor out of wfi
 - wfi in debug mode should execute as a nop

Control whether misaligned accesses are supported via command line