test-dynamo(0711): add dynamo causal-lm test infra + fix gemma3 and qwen3_moe subfunction export

[vbaddi]

Adds tests/dynamo/causal_lm/ — a test suite for the use_dynamo=True export path (PR #1169). Covers 32 architectures across 16 feature columns: export, FP16/BF16/MXFP6/MXINT8 compile, CPU parity (HF PT == QEff PT == ORT), HW parity (ORT == QAIC FP16), CB, blocking-KV, sampler, prefix caching, and multi-device (CCL). HTML + JSON report generated on every run.

Parallelism model:

• Export/compile/CPU-parity lanes are CPU-bound and run fully parallel under pytest-xdist -n 4. Each test writes to a blake2b-keyed workdir (arch + feature + precision hash) so workers never collide and cross-run compile cache reuse works automatically.
• Runtime lanes (QAIC generate, CB, blocking, sampler, prefix caching) carry @pytest.mark.xdist_group(name="qaic-runtime"), which xdist places on a single worker — serialising them. Each test additionally acquires a device from a flock-based DevicePool before touching hardware, so external concurrency (e.g. two pytest sessions running simultaneously) is also safe. Multi-device tests atomically acquire num_devices and release all holds on failure to prevent deadlock.

Two model fixes required for the subfunction lane to pass:

• gemma3: switch RMSNorm to select_interface and bridge cache via QEffSlidingWindowCache.from_legacy_cache.
• qwen3_moe: clone split expert params to remove aliasing in QEffQwen3MoeExperts.__qeff_init__.