Uncaught dnnl::error terminates the process when oneDNN cannot serve a dtype

[lwandrebeck]

Summary

Every oneDNN entry point can throw dnnl::error, and primitive-descriptor
creation does so when the host ISA cannot serve the requested dtype — BF16 without
AVX-512 or AVX2-VNNI-2, for instance. Nothing in matmul_onednn_wrapper() catches
it, so the exception reaches std::terminate and takes the process down.

Reproduce

On a host without AVX-512:

benchdnn --op=matmul --lowoha=true --m=512 --k=512 --n=512 \
         --sdt=bf16 --wdt=bf16 --ddt=bf16 --kernel_name=onednn

terminate called after throwing an instance of 'dnnl::error'
  what():  could not create a primitive descriptor for the matmul primitive.
Aborted (core dumped)          # exit 134

Suggested fix

Catch inside matmul_onednn_wrapper() rather than at the call sites: three of the
five callers (bmm_kernel, bmm_looper, matmul_partitioner) invoke it from
inside an OpenMP parallel region, where an escaping exception may not cross the
region boundary. The AOCL-DLP tail in lowoha_matmul.cpp already documents the
same reasoning for run_dlp.

On failure, mark the AOCL-DLP fallback — the convention the dispatch already uses
for "did not compute" — and let the call sites fall through rather than returning
success on an untouched C.

Reference branch: https://github.com/lwandrebeck/BullDNN/tree/upstream/onednn-uncaught-error-abort