test: add reusable per-architecture XLA validation harness

[inureyes]

Summary

Adds a reusable, two-tier validation harness so adding an OpenXLA (mlxcel-xla) architecture is turnkey, replacing the ad-hoc per-arch flow (dequant a checkpoint offline, run an HF fp32 oracle, run xla_oracle_check by hand). Downstream families (#497-#501) consume it.

What changed

• src/lib/mlxcel-xla/src/validation.rs (new): the structural tier, a pure-Rust byte-exact gate. A per-architecture ArchFixture registry plus check_arch parses a checkpoint's config.json, emits each graph, and diffs it against the committed golden assets/<arch>/*.mlir, localizing the first differing line on drift. emit_graphs is the golden-less freeze primitive for a brand-new family, and MLXCEL_FREEZE_GOLDENS=1 cargo test ... freeze_goldens re-freezes a registered one. It honors MLXCEL_XLA_PRECISION: the goldens are the default f32 graphs, so a byte-exact run under f16/bf16 is rejected with a clear message instead of a confusing diff.
• src/lib/mlxcel-xla/src/lib.rs, src/lib/mlxcel-xla/src/emitter/mod.rs: wire in the validation module (same cfg(any(feature = "iree", test)) gating as the sibling modules), re-export emit_decode_batched, and route the emitter's existing from_json_reproduces_bundled_assets_byte_for_byte test through check_arch so the goldens have a single source of truth.
• scripts/xla/validate_arch.sh (new): the execution tier one command. Given --model <checkpoint> it runs the structural pre-gate, produces the HF fp32 oracle, then runs xla_oracle_check (single-seq greedy == HF oracle) and xla_batch_bench (each batched request == its single-seq reference), reporting pass/fail for both.
• spike/openxla/oracle_continuation.py (new): the arch-generic oracle producer. Loads a checkpoint in fp32, dequantizing an MLX 4bit/8bit checkpoint offline first with the same affine formula as src/weights.rs (--selftest verifies the dequant math against the Rust hand examples), then records the no-EOS greedy continuation.
• src/lib/mlxcel-xla/README.md: documents both tiers and the turnkey add-a-family workflow.

Test plan

• cargo test -p mlxcel-xla --lib validation (structural gate): 6/6 pass; registered_fixtures_are_byte_exact reports a clean byte-exact pass on llama-3.2-1b.
• cargo test -p mlxcel-xla --lib emitter: 14/14 pass, including the now-delegating from_json_reproduces_bundled_assets_byte_for_byte and the rebased refactor: share per-layer attention core across emitter graph kinds #554 gemma2 test.
• cargo clippy -p mlxcel-xla --lib --tests -- -D warnings: clean.
• MLXCEL_FREEZE_GOLDENS=1 cargo test ... freeze_goldens leaves assets/ byte-for-byte unchanged (freeze reproduces the committed goldens).
• oracle_continuation.py --selftest: dequant matches weights.rs (4-bit and 8-bit hand examples).
• Execution-tier live run (scripts/xla/validate_arch.sh --model <ckpt>) needs a real xla-iree build plus IREE GPU/CPU, so it is the opt-in run for a GPU session (not CI), by the two-tier design.

Closes #496