benchdnn matmul silently ignores --kernel_name (always runs AOCL-DLP)

[lwandrebeck]

Summary

The LOWOHA matmul harness never applies --kernel_name. Two defects combine so
every value runs AOCL-DLP:

1. VALID_KERNEL_NAMES (benchdnn_utils.hpp) omits native_gemm,
   native_brgemm and the auto_tuner spelling of auto — all real
   matmul_algo_t values. An unlisted name is not an error: matmul_utils.cpp
   silently substitutes aocl_dlp / aocl_dlp_blocked.
2. set_lowoha_matmul_params() populates params.dtypes but never
   params.lowoha_algo, which is the field kernel_select() reads. It stays
   matmul_algo_t::none, and kernel_select() turns that into
   aocl_dlp_blocked.

On a --no-aocldlp build every --kernel_name therefore dies after ~2 ms with
"Selected kernel aocl_dlp_blocked requires AOCL-DLP", making matmul
unbenchmarkable on any host that cannot use AOCL-DLP. On a build that does have
AOCL-DLP the failure is quieter and arguably worse: the reported numbers are
AOCL-DLP's regardless of which algorithm was requested.

The library itself is fine — the ops-API path already honours the name via
set_forced_kernel().

Separate observation

The matmul driver performs no output comparison at all, so it cannot detect a
wrong result — only an execution failure. A kernel that computes garbage quickly
passes benchdnn. That made it unusable for validating the kernel work behind these
reports; the gtests were needed instead.

Suggested fix

Add the missing names and assign params.lowoha_algo from the requested name.
Unknown and "none" should map to matmul_algo_t::none so the library's own
selection still applies, which keeps ZENDNNL_MATMUL_ALGO round-tripping.

Reference branch: https://github.com/lwandrebeck/BullDNN/tree/upstream/benchdnn-kernel-name-ignored