Stack from ghstack (oldest at bottom):

• [executorch][vulkan] Fetch embedding_q4gsw indices of rank 2 and 3 from textures #25025
• [executorch][native] Keep float scales of int4 ggml embeddings exact #25024
• -> [executorch][native] Record embedding op samples #25023
• [executorch][native] Quantize nn.Embedding weights in the op sample schemes #25022

Embedding becomes the second op in the op samples, with a case module like linear's: the backend test suite's embedding models, LLM vocabulary and token shapes, leading batch dims, a scalar index, the first and last rows, int32 indices, padding_idx, one table per GGUF super-scale family, and dynamic token and batch dims. It fits the float, int4 weight-only and GGUF schemes.

draws and chain take a sampler, standard normal by default, so a case can draw integer indices. Parallel moves to cases.py for both ops, and its modules are now branches, which renames the constants of linear's scale_families packages; their values are unchanged. Only weight-only packages with an nn.Linear in a model that fits the scheme get golden-int8_act, because without one it would repeat golden; Scheme.adds_int8_act decides this from the modules int8_activations quantizes.

The native test workflow (test-native.yml) and the CMake sample rule now also depend on the backend test suite's test_embedding.py, so a change there rebuilds and retests the samples, as test_linear.py already does.

A new ggml-CPU sample test, ggml_sample_test_no_linear in Buck and native_ggml_sample_test_no_linear in CTest and the OSS ggml job, judges every op but linear without activation quantization, so CI also covers embedding on plain ggml-CPU. Linear keeps only the act_quant test. Until the next diff, both CPU CI targets exclude the 22 int4 weight-only f32 embedding tests that ggml fails.

Authored with Claude Code.

Differential Revision: D130869556

NOTE FOR REVIEWERS: This PR has internal Meta-specific changes or comments, please review them on Phabricator!