Everything else should already be handled.

Since we had to rewrite the deref walking loop anyway, I took the
opportunity to make it a bit clearer and more efficient.  In particular,
in the AoA case, we will now emit one minmax instead of one per array
level.

This commit removes most of the deref instruction lowering.  Instead of
lowering early, we only lower textures and images and we only do so
right before any of the anv image lowering passes.

This commit completely reworks function calls in NIR.  Instead of having
a set of variables for the parameters and return value, nir_call_instr
now has simply has a number of sources which get mapped to load_param
intrinsics inside the functions.  It's up to the client API to build an
ABI on top of that.  In SPIR-V, out parameters are handled by passing
the result of a deref through as an SSA value and storing to it.

This virtue of this approach can be seen by how much it allows us to
delete from core NIR.  In particular, nir_inline_functions gets halved
and goes from a fairly difficult pass to understand in detail to almost
trivial.  It also simplifies spirv_to_nir somewhat because NIR functions
never were a good fit for SPIR-V.

Unfortunately, there is no good way to do this without a mega-commit.
Core NIR and SPIR-V have to be changed at the same time.  This also
requires changes to anv and radv because nir_inline_functions couldn't
handle deref instructions before this change and can't work without them
after this change.

We were only initializing vtn_builder::func for the pre-walk where we
build the CFG.  We were only initializing the nir_builder for the later
walk through the instructions even though were were setting b->cursor
for the pre-walk.  Let's set both both places so that everything is
consistent.  This useful because we handle OpFunctionParameter in the
pre-walk and we're going to need to be able to emit instructions.

Now that pointers can be derefs and derefs just produce SSA values, we
can convert any pointer to/from SSA.

Previously, pointers fell into two categories: index/offset for UBOs,
SSBOs, etc. and var + access chain for logical pointers.  This commit
adds another logical pointer mode that's deref + access chain.

It's tempting to think that we can just replace variable-based pointers
with deref-based or at least replace the access chain with a deref
chain.  Unfortunately, there are a few sticky bits that prevent this:

 1) We can't return deref-based pointers from OpVariable because those
    opcodes may come outside of a function so there's no place to emit
    the deref instructions.

 2) We can't always use variable-based pointers because we may not
    always know the variable.  (We do now, but he upcoming function
    rework will take that option away.)

 3) We also can't replace the access chain struct with a deref.  Due to
    the re-ordering we do in order to handle loop continues, the derefs
    we would emit as part of OpAccessChain may not dominate their uses.
    We normally fix this up with nir_repair_ssa but that generates phi
    nodes which we don't want in the middle of our deref chains.

All in all, we have no real better option than to support partial access
chains while also re-emitting the deref instructions on the spot.

Now that push constants are using on-the-fly offsets, we no longer need
to handle access chains in vtn_pointer_to_offset.

Push constants have been a weird edge-case for a while in that they have
explitic offsets but we've been internally building access chains for
them.  This mostly works but it means that passing pointers to push
constants through as function arguments is broken.  The easy thing to do
for now is to just treat them like UBOs or SSBOs only without a block
index.  This does loose a bit of information since we no longer have an
accurate access range and any indirect access will look like it could
read the whole block.  Unfortunately, there's not much we can do about
that.  Once NIR derefs get a bit more powerful, we can plumb these
through as derefs and be able to reason about them again.

Before, we were doing structure splitting in spirv_to_nir.
Unfortunately, this doesn't really work when you think about passing
struct pointers into functions.  Doing it later in NIR is a much better
plan.

Once we've gotten rid of everything but the main entrypoint, there's no
reason why we should go ahead and lower them all.  This is what radv
does and it will make future work easier.

This adds a concept of "members" to a variable with an interface type.
It allows you to specify the full variable data for each member of the
interface instead of once for the variable.  We also add a lowering pass
to lower those variables to a sequence of variables and rewrite all the
derefs accordingly.

The only thing still using old-school drefs are function calls.

This pass doesn't handle deref instructions yet.  Making it handle both
legacy derefs and deref instructions would be painful.  Since it's not
important for correctness, just disable it for now.

Sometimes it's useful for a pass to be able to clean up its own derefs
instead of waiting for DCE.  This little helper makes it very easy.

This commit introduces a new nir_deref.h header for helpers that are
less common and really only needed by a few heavy-duty passes.  In this
header is a new struct for representing a full deref path which can be
walked in either direction.