Convert basic `parallel_for` kernel launches to `nd_range`

[kris-rowe]

Details

The work group size chosen by the SYCL runtime for the basic parallel_for kernel launches is suboptimal in some cases. These should be converted to the nd_range version and an appropriate default work group size should be found in each case.

Note: basic parallel_for usage is limited to the sycl-ref backend.

Todo

• Get baseline performance data for relevant kernels
• Evaluate performance using nd_range kernel launch

Vector

• Norm (x3)
• Reciprocal
• Scale
• AXPY
• PointwiseMult

ElementRestriction

• StridedNoTranspose
• OffsetNoTranspose
• StridedTranspose
• OffsetTranspose

Basis

• ApplyWeight
• ApplyNonTensorInterp
• ApplyNonTensorGrad
• ApplyNonTensorWeight

QFunction

• Apply

Operator

• LinearDiagonal
• Linear Assemble