Prototype AMD XDNA bfp16ebs8 as a backend-local vendor tier

[aravishankar-mp]

Part of #110. Blocked by #146.

Goal

Prototype the proven AMD XDNA bfp16ebs8 block-8 contract as a backend-local vendor experiment so
the native hardware path can be exercised before TOSA standardization settles.

This is deliberately not called block FP16, BFP16, BF16, TOSA MX, or EXT-MX-*. It must not be
accepted under the stable TOSA 1.0 artifact format or advertised through
TosaCapabilityProvider.

Boundaries

• Use a backend-local/precompiled experimental artifact with an explicit version and target
  identity that cannot collide with released TOSA.
• Keep the numerical label guest-visible and immutable; a host knob may disable the experiment but
  cannot substitute it under another tier.
• Reuse the characterization ticket's reference model and fixtures exactly.
• Bind caller-owned buffers directly and perform no submission-time repacking or scale synthesis.
• Compilation remains offline/load-time and pinned to the measured toolchain identity.

Acceptance criteria

• The experimental artifact encodes every scale/value plane and block-layout parameter needed to
  reproduce the characterized contract without hidden defaults.
• Exact or explicitly tolerance-bounded oracles cover edge classes and representative block
  exponent disagreement.
• The full lifecycle runs on the reference NPU with direct-binding diagnostics and no hidden
  staging.
• Unsupported or mismatched artifacts are rejected before native resource creation.
• Documentation labels the surface “AMD bfp16ebs8 vendor experiment (block-8)” and explains why
  it is not a TOSA capability.
• The prototype creates no protocol 1.0 wire change and no claim about future MX compatibility.

[aravishankar-mp]

Crate integration is up as draft PR #165 (design in docs/plans/issue-148-bfp16ebs8-vendor-tier.md on PR #161's branch): XBFP container with derived slot plan, immutable-identity load arm reusing the audited executable path, vendored #146 model as a host test target, and an on-metal suite — 12 passing on the reference NPU including the full guest-quantized K=512 MXINT8 MATMUL lifecycle, bit-exact against the fold oracle.

One acceptance criterion is deliberately not met yet: probe P6 falsified the planned per-step-RNE accumulation oracle at tie-adjacent steps (serial chain, no persistent guard bits, but crafted ties break toward zero while an organic mid-chain tie broke away — no single rounding model fits). The order-contract test is quarantined behind an #[ignore] naming the reason, and the PR stays draft until P6 pins the accumulator's real tie rule. Probes: tests/bfp_p6_probe.rs.