feat: add Youtu-VL XLA vision and MLA inference (#872)

[inureyes]

Summary

• Add an explicit Youtu-VL XLA text configuration for the checkpoint's dense
  DeepSeek-style MLA, compressed KV layout, and interleaved RoPE.
• Add bounded IREE vision buckets with checked grid planning, 2D vision RoPE,
  window/full attention selection, output restoration, and the built-in merger.
• Match the pinned HF processor's spatial-merge-grouped, channel-fast flattened
  patch layout and expand one logical placeholder per image to the exact merged
  token count.
• Recover tensors from stale safetensors index hints by scanning every named
  shard while rejecting missing, duplicate, and mixed namespaces.
• Preserve explicit BF16 vision carrier boundaries while reusing the shared
  numeric LayerNorm, RMSNorm, GELU, and softmax decompositions.
• Add an immutable independent HF-eager oracle pinned to
  tencent/Youtu-VL-4B-Instruct@8d30a0e49662a1d628a472b12df264dbcd768753
  and a fail-fast production runner that writes partial diagnostic tensors
  before stopping at the first unchanged-threshold failure.
• Keep the family capability fail-closed while intermediate and token-exact
  production gates remain unresolved.

Validation

• All 18 pinned checkpoint artifact hashes and the fixture hash matched.
• cargo test --features xla-iree youtu_vl_ --lib: 12 passed, 1 ignored
  (the ignored gate requires the production checkpoint/runtime environment).
• Focused HF patch-layout and stale-index resolver regressions passed.
• python -m unittest test_youtu_vl_reference_oracle.py: 13 passed.
• cargo fmt --all --check
• cargo clippy --fix --allow-dirty --features xla-diagnostics --example xla_youtu_vl_reference_check
• git diff --check
• The pinned HF capture and local-task IREE run completed with bounded
  diagnostics-only worker topology.

Actual oracle evidence

• Resized pixels, flattened patches, window ordering, spatial metadata, and
  vision RoPE match the pinned HF capture.
• Patch projection passes the unchanged atol=0.02, rtol=0.02 gate
  (max_absolute=0.015625).
• Layer 0 passes that gate (max_absolute=0.03125 with relative tolerance).
• The first failing selected stage is layer.7.full: flat index 310,
  actual 0.3125, reference 0.28515625, absolute 0.02734375,
  max_absolute=0.203125, and 2,999 unchanged-threshold mismatches.
• Feeding the IREE patch projection into the HF layer-0 implementation stays
  within 0.015625 of the HF reference, while the IREE layer-0 output reaches
  0.03125. The row-wise LayerNorm decomposition matches HF exactly; the first
  drift is the backend BF16 projection/attention contraction schedule.
• Explicit BF16 contraction input casts produced byte-identical IREE results,
  so that ad hoc workaround was removed. No tolerance was changed.
• The fail-fast runner now stops this case in roughly 39 seconds after build and
  retains every selected vision tensor for analysis instead of spending the
  prior full-run duration on already-invalid language output.

Draft limitation

The current local CPU IREE target cannot establish production-reference BF16
contraction parity for the pinned HF eager schedule. Per #932, this is recorded
as an unresolved backend numeric contract, not mathematical equivalence.
Language logits/KV, greedy tokens, reset/reuse, and mixed continuous batching
therefore remain unqualified. This PR stays draft and the capability stays
fail-closed until the unchanged intermediate and token-exact gates pass on a
reproducible production-relevant contract.

Closes #872

[inureyes]

Rebased onto main@7fe1412d after #934 and force-pushed as 7886a657.

Post-rebase validation:

• cargo fmt --all -- --check
• git diff --check
• cargo test --lib youtu_vl: 24 passed
• cargo test -p mlxcel-xla --features iree --lib youtu_vl: 9 passed, 2 intentionally ignored
• real 27-layer actual-schedule IREE CPU compile gate: 1 passed

The runtime now uses the explicit new_legacy_unqualified auxiliary-artifact path. That is intentional: #934 is only the versioned numeric-contract foundation, while the complete pinned ~10 GB HF-reference-versus-production numeric report is still pending. This PR therefore remains draft and must not be merged yet; #872 remains blocked by #932.

[inureyes]

Published the bounded Youtu-VL follow-up at 20c2bffd:

• exact logical placeholder expansion with fail-closed partial-expansion checks;
• stale safetensors index recovery by scanning every shard named by the index;
• pinned HF spatial-merge/channel-fast patch layout and corrected oracle reconstruction;
• fail-fast preprocessing/selected-vision comparisons with partial diagnostic artifacts;
• explicit BF16 vision carrier boundaries using the shared numeric helpers.

Validation:

• Rust Youtu suite: 12 passed, 1 environment-gated ignored
• Python oracle contract: 13 passed
• focused patch-layout, shard-resolution, and emitter tests passed
• fmt, clippy, and git diff --check passed

The real pinned checkpoint now reaches the unchanged intermediate gate quickly.
Preprocessing and metadata are exact; patch projection passes
(max_absolute=0.015625), and layer 0 passes. The first selected failure is
layer.7.full at flat index 310: actual 0.3125, HF 0.28515625,
absolute=0.02734375, max_absolute=0.203125, 2,999 threshold mismatches.

An explicit BF16 contraction-input experiment produced byte-identical IREE
results and was removed. The LayerNorm decomposition matches HF exactly; the
remaining drift begins at the backend BF16 projection/attention contraction
schedule. No tolerance changed. The PR remains draft and fail-closed; language
KV/logits/tokens and lifecycle gates are not claimed.