[compiler] Support cl_intel_required_subgroup_size

Note: this commit does not yet enable the extension by default! See
below.

This commit adds support for the cl_intel_required_subgroup_size OpenCL
extension.

* It supports device-specific lists of sub-group sizes by adding two new
  fields to the Mux device info available to targets: a pointer to a
  list of sub-group sizes, and a size for the list. The pointer should
  point to static/managed data.
* It supports the kernel data necessary for the extension by extending
  `compiler::KernelInfo` with the (optional) required sub-group size
  and the 'spill mem size'. The required sub-group size is
  encoded/decoded through the binary header. The spill size is not, but
  that's in keeping with the private mem size - this might indicate an
  oversight for offline-compiled binaries.

The rest of the work is plumbing that data into the extension's queries,
and adding testing coverage.

Two new compiler passes have been added to force the compiler into
failing if the required sub-group size is not supported by the device,
or if the requirement has not been satisfied by the compiler. Both of
these are always true as none of our targets advertise any sub-group
sizes, and never force the vectorizer into vectorizing to that size.

The `muxc` compiler tool can be given supported sub-group sizes via a
command-line option to ensure these compiler passes can be tested in a
device-agnostic way.

Note that this doesn't include support for this extension when the
program was compiled with SPIR-V (clCreateProgramWithILKHR). In a strict
reading of the extension spec, `CL_KERNEL_COMPILE_SUB_GROUP_SIZE_INTEL`
only reports the value of the attribute and thus has no meaning for
SPIR-V, but we can be charitable here.

It is expected that this property will manifest itself via the
`SubgroupSize` execution mode, but more work needs to be done here. For
that reason the extension is left disabled until SPIR-V support is
added.

Author: frasercrmck

llvm/lib/SYCLNativeCPUUtils/compiler_passes/compiler_pipeline/docs/utils.rst

@@ -1067,6 +1067,23 @@ the ``NoUnwind`` attribute can still be generated in certain cases. This pass
 adds the ``NoUnwind`` attribute to every function in the module, for target code
 generators that can't handle exceptions.
 
+VerifyReqdSubGroupSizeLegalPass & VerifyReqdSubGroupSizeSatisfiedPass
+---------------------------------------------------------------------
+
+These passes check whether the compiler can handle, and has successfully
+handled, a kernel with a required sub-group size.
+
+The ``VerifyReqdSubGroupSizeLegalPass`` searches for any kernel with a required
+sub-group size and checks whether the device supports such a size. It does this
+using the target's ``compiler::utils::DeviceInfo`` analysis. Any unsupported
+size results in a compiler diagnostic, which the compiler can handle (usually
+via a build error).
+
+The ``VerifyReqdSubGroupSizeSatisfiedPass`` searches for any kernel entry point
+with a required sub-group size and checks whether the vectorizer was able to
+satisfy that requirement. As such, it should be run after vectorization. A
+compiler diagnostic is raised for each kernel for which this does not hold.
+
 Metadata Utilities
 ------------------