Decode BF16 weights and score merged MoE expert tensors

[SloptimistPrime]

The imatrix probe cast stored BF16 bytes to floats instead of decoding the weights. Decode BF16 correctly and handle merged MoE tensors with separate input-channel importance for each expert.

Record MoE routers separately and require them to retain their source precision. Unknown tensor families, invalid shapes and non-finite costs still stop the probe.

22 tests pass using real GGUF files and direct cost calculations. On Qwen3-30B-A3B, 26 sampled matrices containing 66,060,288 values decode exactly against the original safetensors. A small HF-versus-GGUF prediction check passed preselected limits in all three domains.

Includes public dataset revisions, row indices and hashes for reproduction. Expanded calibration reduced partial-expert warnings from 36 to 28, but coverage is still incomplete. These scores estimate rounding error; they are not measured quality gains or a completed allocation-versus-uniform comparison.

Results and limits: notes/imatrix-probe-gb10.md.

I used AI assistance and reviewed and tested the changes.