[ARM64] Add base bf16 mmt4d tiles for cores without the BF16 extension (#24900)

On Arm64 cores without the Armv8.6 BF16 extension, bf16 `mmt4d`
currently falls back to the scalar generic kernel. All existing Arm64
bf16 tiles require the `_bf16` feature:

```text
IREE_UK_MMT4D_TILE(arm_64, bf16, bf16, f32,  1..8, 8, 4, _bf16)
IREE_UK_MMT4D_TILE(arm_64, bf16, bf16, bf16, 1..8, 8, 4, _bf16)
```

This PR adds base Arm64 bf16 tiles for cores without that extension.

The new tiles widen bf16 operands to f32 using NEON, accumulate with
FMLA, and narrow back to bf16 on store when required. The same
implementation serves both bf16 and f32 accumulator/output variants.

No compiler changes are required: `enumerateMatmulTileArm64` already
selects `8x8x1` tiles for bf16 when the BF16 extension is unavailable.

## Performance

i.MX95 (Cortex-A55, six cores), 512x512x512, median of three runs:

| case | before | after | speedup |
|---|---:|---:|---:|
| bf16 -> bf16 | 216 ms | 7.64 ms | 28x |
| bf16 -> f32 | 216 ms | 7.53 ms | 29x |
| f32 -> f32 | — | 8.22 ms | — |

For narrow M, using Mx512 by 512x512:

| M | before | after | speedup |
|---|---:|---:|---:|
| 1 | 1.52 ms | 0.879 ms | 1.7x |
| 2 | 2.31 ms | 0.874 ms | 2.6x |
| 4 | 3.37 ms | 0.887 ms | 3.8x |

## Correctness

Tested on device with `iree-run-module` for M=1,2,4,8 for bf16 -> bf16
and M=8 for bf16 -> f32.

The registered tiles match the generic implementation for normal inputs.
The widening behavior for bf16 subnormals and NaN payloads follows the
native NEON conversion behavior and can differ from the scalar helper,
similarly to the existing Arm64 f16 tiles.

The bf16 -> bf16 path accumulates in f32 and rounds once when storing,
matching the behavior of the existing Arm64 f16 tile.

Assisted by: Claude

---------

Signed-off-by: Thomas Ziereis <ziereis@roofline.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>