Add optimizations to collision kernel for AMD GPUs

[gsitaram]

Change block size and use template parameter instead of _NUMVECTORS in GPU_CollideStream_mMidFluidCollision_mWallCollision_sBB_WallShearStress kernel to improve achieved memory bandwidth in the kernel. The larger block size helps lower TLB misses and the template parameter makes the number of vectors argument a compile time constant that helps the compiler avoid spilling local arrays to slower scratch memory.

Please test and accept these changes to significantly improve HemeLB performance on AMD GPUs.