Merge branch 'master' into cycles_bvh

Cycles BVH: Add some TODOs and some minor cleanup

Cycles BVH: Lazily allocate aligned transform in BVH node

This is an attempt to save up memory for oriented nodes during BVH build.

Will likely introduce some slowdown to construct hair BVH, but that we can
optimize later by using specialized allocator.

Cycles BVH: Bring back extmax to regular BVH aligned node intersection on GPU

Also fix compilation error there.

Cycles BVH: Don't store w component of transform for regular bvh unaligned nodes

Saves quite some memory (467MB vs 528MB with koro.blend file).

Once again addressing review from Brecht.

Cycles BVH: Simplify some evaluation in regular BVH traversal

Addressing review from Brecht.

Cycles BVH: Simplify fetching of regular BVH unaligned transform

Nice trick suggested by Brecht, thanks!

Cycles BVH: Use proper forward eclaration of Transform

Spotted by Brecht, thanks!

Cycles BVH: Avoid confusion about shadowing class members

Cycles BVH: Fix typo

Cycles BVH: Move unaligned threshold to parameters

Suggested by Brecht in review.

Should be no functional changes.

Cycles BVH: Remove debug-only code

That code would have created too much unaligned nodes.

Spotted by Brecht, thanks!

Merge branch 'master' into cycles_bvh

This commit also removes NO_EXTENDED_PRECISION from refactored code, solving
merge conflicts and getting rid of this define in new code.

Cycles: Avoid unwanted evaluation in non-hair traversal

Merge branch 'master' into cycles_bvh

Cycles: Bring back extmax to aligned node rebust intersector

Hopefully will solve any possible regressions in behavior.
However, all unit tests are passing just fine.

Cycles: Fix wrong order of node_address and visibility arguments

Cycles: Finish optimization from previous commit, now includes non-SSE2 kernels

Cycles: Optimization for non-hair regular BVH traversal

Cycles: Hook up intersectors to non-SSE2 BVH traversal

Seems to be passing regression tests on CPU. CUDA kernels also seems to work fine.