[spirv] Gate NonPrivatePointer on cooperative matrix loads

SubgroupMatrixLoad() hardcodes SpvMemoryAccessNonPrivatePointerMask on
OpCooperativeMatrixLoadKHR. Ordinary accesses do not: the printer's
MemoryAccessMaskForPointer() requests it only under the Vulkan memory
model, for storage or workgroup pointers, and only when the access mode
is read_write. The same file already follows that rule for storage
textures, via NonPrivateTexel.

NonPrivatePointer makes the access take part in the memory model's
availability and visibility rules. A read-only pointer has nothing to
take part in, since no agent can write it for the life of the dispatch,
but honouring the operand still keeps the access out of storage that
other agents cannot observe.

Apply the same rule. Only its access mode term needs testing here: these
builtins accept nothing but storage and workgroup pointers, and subgroup
matrix already requires the memory model. SubgroupMatrixStore() is left
alone: storing needs a writable pointer, so the rule could never change
its result.

Measured with two builds differing only in this change, on Phi-4-mini
prefill at 512 tokens through ONNX Runtime's subgroup-matrix matmul
kernel, which reads its A operand from a var<storage, read> buffer: an
AMD Radeon RX 7900 XTX goes from 2300 to 5058 tok/s. On an NVIDIA
GeForce RTX 5080 and an Intel Arc 140V there was neither improvement nor
regression.

The added test covers a read-only pointer; the existing subgroup matrix
tests all use read_write, whose result is unchanged.

Bug: none
Change-Id: Iea60136b3aabe3274141bf1cdb93db2d7d4879a0
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/330695
Reviewed-by: Alan Baker <alanbaker@google.com>
SLSA-Policy-Verified: SLSA Policy Verification Service <devtools-gerritcodereview-exitgate@google.com>
Commit-Queue: Yang Gu <ygu@microsoft.com>