Add intrinsic for efficient matrix load/store/prefetch

The LSC block 2D operations require a matrix descriptor as an argument.
When the vISA finalizer creates the descriptor, it cannot optimize loop-
invariant instructions. So, the finalizer generates a lot of redundant
MOV instructions. This patch adds intrinsics for efficient matrix load,
store, and prefetch operations. The intrinsics take the user-defined
matrix descriptor, which can be properly optimized.

 Add an intrinsic for named barrier arrive/signal operation

.

 Fixed warning from static code analyser

.

 Add support for XeHPCVG platform

.

 Add support for Arrow Lake and Lunar Lake devices (#201)

.

Co-authored-by: Victor Mustya <victor.mustya@intel.com>

 Prevent creation of scalable vectors

To create a vector IITDescriptor::getVector() should be called with isScalable arg explicitly set to false

 Mark new lsc_load_merge intrinsics as LSC ones

.

Allow kernels to be referenced in @llvm.global.annotations

---------------------------

Handle circular dependencies between structs

---------------------------

LLVM 15 build fix

.