Cycles: Fix rendering of barbershop scene

Merge branch 'master' into split-kernel-faster-building

Cycles: Fix issue with AO in split kernel_path_background

Watch out for changes to variables passed by value, these changes
don't persist across the split.

Cycles: Fix build with OSL enabled

Cycles: Split shader_eval out of the kernel_lamp_emission kernel

With this kernels for BWM and classroom scenes are building in half the
time as master. Render times are 1% faster as well.

Cycles: Split indirect_lamp_emission

Cycles: Split kernel_path_lamp_emission

Cycles: Only use first lamp in indirect_lamp_emission

Produces different results when lamps overlap from POV of a ray, but allows
for this function to be split. The differences probably wont be noticeable
in most scenes. Old behavior could be reattained by placing lamps
into the BVH.

Cycles: Fix mistake with ShaderEvalTask that could lead to crash

Cycles: Merge shadow_blocked_ao and _dl kernels to improve build times

Gives another 4 seconds improvement.

Cycles: Remove early out that may cause threads to diverge before barrier

Cycles: Use ShaderEvalIntent more to pass flag, closures, and eval_task less

This makes it so that path_flag and max_closures are not passed to shader eval
functions, instead a limited number of ShaderEvalIntents are used. This also
removes the need for ShaderEvalTask in the mega kernel and simplifies the
code a bit.

Cycles: Modify kernel_indirect_background to remove call to svm_eval_nodes

Speeds up kernel building by another 2-3 seconds. Rendering is
another 1% slower.

Cycles: Add shader_eval_state to allow eval of ray states besides RAY_ACTIVE

Cycles: Split kernel_path_background

Cycles: Split the indirect_background function

Cycles: Fix shadow catcher and branched path after recent split changes

Cycles: Split apart direct_emission around shader_eval

Cycles: Speed up split kernel builds by using split direct_emission functions

This removes the indirect call to `svm_eval_nodes` from the
`kernel_direct_lighting` kernel, and speeds up kernel build times
by 4-5 seconds. Render time is slower by ~1%. It should be possible
to apply this technique to other areas of the kernel to get even
better build times.

Note: this breaks branched path and shadow catcher in the split
kernel. Should be an easy fix, will do later.

Cycles: Split direct_emissive_eval around shader_eval