Add callbacks on functions that really need them instead.

- alignment is set on each single IncSP now instead of a global arch_env
  setting
- fixed ia32 not requesting any alignment anymore (even on darwin where
  we need it)
- fixed amd64 creating IncSP nodes without flag_modifies_flags set

We now add missing keeps exactly before register allocation where we
actually need them (so that unused outputs which still need registers
assigned are not missed by the register allocation).

This removes functions from the public API which are pointless outside
localopt rules, and also avoids another struct in the public API.

- It was only available in the ia32 backend and broken there. As the
  feature is rarely used, I remove it for now instead of going through
  the trouble of fixing it.

The previous approach put the irg link into the block attributes, so
while saving memory an extra indirection was necessary. Some testing
with valgrind massif+callgrind revealed that a direct link increses peak
memory usage by ~1% while also increasing runtime by ~1%. More
importantly the code gets a bit simpler and some strange corner cases can be
removed with the direct link.

Implemented for now: When the flags of one Sub or Cmp node are needed, but
the flags of another one are currently alive, invert the condition code and
avoid rematerialization.

At leat that's what qemu and arm-linux-gnueabi assume by default
TODO: add a commandline switch for endianess

- We now create a spill/reload/phi-web and take the largest necessary
  reload size to determine the size of the entities and the spills.
- Some cleanups

This also disables the memcombine optimisation if unaligned stores are
not supported by the backend.
We may add some tests in the future to reenable this in cases where we
can prove that the combined store is aligned.