cpu: aarch64: add f16 and bf16 support to jit_io_helper and jit_uni_binary

[Sqvid]

Description

This PR adds f16 and bf16 support to jit_io_helper_t. I have also demonstrated the impact of these changes by adding support for these data-types to jit_uni_binary_t.

# Dispatch before:
$ ./build/tests/benchdnn/benchdnn --binary --batch=test_binary_{float16,bfloat16} --batch=harness_binary_{f16,bf16}
...
============================================================
= Implementation statistics (--summary=no-impl to disable) =
============================================================
| ref:any : 8900 (100%)                                    |
============================================================
tests:20000 passed:8872 skipped:11100 mistrusted:28 unimplemented:0 invalid_arguments:0 failed:0 listed:0

# Dispatch after:
$ ./build/tests/benchdnn/benchdnn --binary --batch=test_binary_{float16,bfloat16} --batch=harness_binary_{f16,bf16}
...
============================================================
= Implementation statistics (--summary=no-impl to disable) =
============================================================
| ref:any : 6240 (54%)                                     |
| jit:uni : 5320 (46%)                                     |
============================================================
tests:20000 passed:11520 skipped:8440 mistrusted:40 unimplemented:0 invalid_arguments:0 failed:0 listed:0

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

[michalowski-arm]

LGTM

[jondea]

How do we know we can always use z31? Also, I think a comment explaining that this is a fallback would be useful.

[jondea]

Can we get rid of this mov with

host_->lsl(dst_vmm.s, src_vmm.s, 16);

If not, a comment explaining why would be useful