Use early output check in bsdf_alloc()

Indeed seems better approach than storing CLOSURE_NONE_ID.
Initially, missed that it's possible to ignore those with a simple
early output though :)

Fixes while chasing regression suit failure of the previous
version of this patch.

Use more NaN/inf safe way to check sample weight. It is not
fully robust, hence I did not add comment about it. For now
what is the most important is avoid functional changes for
the regular files. So from this point of view used the same
way of comparison as it used to be.

Ideally would add an assert in the BSDF allocation to catch
cases when non-finite weight is passed to catch actual root
causes.

The failing test was tangent_missing_uv.blend. I'll give it
a deeper dig outside of this patch to track where non-finite
BSDF weight is coming from.

Additionally, fixed possible access to the first closure if
the shader has no closures. This is a bit annoying, but seems
to be right thing to do, and, hopefully, will not show up in
a profiler.

Added a brief comment about using the if the way we use it
(so that people pay a bit more attention when/if converting
to an early output).

Removed changes in the phase sampling and BSDF picking.

BSDF picking is only used when shader data is tagged with presense of
BSDF, so it must have at least one BSDF. So the check was redundant.

The volume is a bit more tricky to me to fully follow, but also seems
that it is only used from place where we know we have volume.

In any case, lets play it safe and commit allocation changes which
we know work and solve actual problem. The rest we can re-iterate
separately.

I'll commit the allocation fix to 2.93, This is what I understood you
prefer as well :)