'A' is for 5-bit immediates, but the code matches the behavior of 'i', i.e. any immediate.

Instead of having a generaic main loop in be_main with several callbacks
into the targets, split the code in be_main up into a set of convenience
functions and let the targets implement the main loop itself. This is
nicer because:

- The control flow into and out of the functions in the target becomes
  obvious as no external party is using callbacks anymore.
- Targets can insert custom code at any place between the generic
  passes.
- In the future we can parameterize the generic backend passes directly
  instead of stuffing info into generic isa_if structures.

Matze's cleanup broke things 8ebe31d226843d6f2cdfcc1ee100ca607c79ba65
Apparently, value==result shall be explicitly allowed, but in the case
of NaN, result was modified and afterwards a read from value returns the
wrong results. Fix was to move the reads before the writes.

Also, a cast from implicit one to explicit one forgot to set the one.
Likewise, the size comparison was oblivious of explicit ones.

Adds few more asserts.

Adds a unittest to check NaN payload stuff.

The optimization was disabled by 8e35902f.

It is only passed around recursively.

This makes the code a bit clearer.

This fixes opt/minus_zero.c.

The idea here, is that you do not simply set a target operating system
or architecture, but have a setting for all things different between the
operating systems. Knowledge about the name and habits of the operating
system is left to the frontend.
(We may want to provide a small support library in the future to share
the triple parsing and OS -> settings code between multiple frontends).