feat: add Qwen3-MoE and OLMoE on the shared MoE FFN primitive

[inureyes]

Summary

Extend the mixture-of-experts family coverage (issue #501) on top of the shared MoE FFN primitive (#500) and the qk-norm attention core (#497). Qwen3-MoE and OLMoE both route with the exact softmax-over-all-experts, top-k, norm_topk_prob renormalization the primitive already emits and have no shared expert, so they land as a config plus weight-naming delta on the existing emit. Families whose routing or attention the shared core does not provide are deferred with a specific message rather than mis-emitted.

What changed

• emitter/config.rs: Config::from_json now parses qwen3_moe (per-head q/k RMSNorm attention, no bias, experts on moe_intermediate_size) and olmoe (flat q/k RMSNorm on the standard pre-norm block, experts on intermediate_size, clip_qkv rejected as a follow-up), each building a MoeConfig with no shared expert. DeepSeek-V2/V3, PhiMoE, GLM4-MoE, dots1, ERNIE-4.5-MoE, and gpt-oss are rejected with a blocker-specific message.
• weights.rs: a weight_specs test locks the new families' expert-bank naming on the mlp prefix (mlp.gate.weight, mlp.switch_mlp.{gate,up,down}_proj.weight), matching the mlx-lm switch_mlp checkpoint layout, with no shared-expert / dense-MLP / bias tensors. The existing generic MoE weight_specs needed no change.
• emitter/mod.rs: replaced the outdated qwen3_moe deferral test with parse tests for the two new families, an emit test that the qk-norm attention composes with the MoE FFN across every graph kind, and a deferral test naming each blocked family's reason.
• validation.rs: registered byte-exact structural fixtures qwen3-moe-tiny and olmoe-tiny (per-head vs flat q/k-norm shapes locked, router + stacked-expert args present, no shared/dense-MLP args); generalized the golden-freeze helper over the MoE fixtures.
• spike/openxla/moe_oracle.py: block-level checks for Qwen3-MoE and OLMoE against an HF fp32 MoE block.

Validation

• Block-level token-exact vs HF fp32 (spike/openxla/moe_oracle.py, IREE llvm-cpu): Qwen3-MoE max block diff ~1.9e-9, OLMoE ~2.2e-8; Qwen2-MoE / Mixtral still pass.
• cargo test -p mlxcel-xla --lib: all tests pass (byte-exact gate now covers both new fixtures; all existing dense + MoE fixtures byte-identical).
• cargo clippy -p mlxcel-xla --lib --tests -- -D warnings: clean.
• Full-model token-exact and xla_batch_bench serve reference-exact need the xla-iree build and are deferred to the post-merge gate (block-level + structural landed in-agent, allowed by the epic).

Test plan

• spike/openxla/moe_oracle.py block-level oracle (all four MoE families PASS)
• cargo test -p mlxcel-xla --lib (all tests pass, byte-exact gate green)
• cargo clippy -p mlxcel-xla --lib --tests -- -D warnings (clean)
• existing dense + MoE fixtures byte-identical (no golden drift)

Closes #501