[executorch][native] Keep float scales of int4 ggml embeddings exact - #25049

Pull Request resolved: #24975

An int4 weight-only embedding stores its group scales in the PTN as f32 or bf16, and eager dequantizes with those exact values. ggml's Q4_0 blocks hold only an fp16 scale, so the ggml engine rounded each scale to fp16 at load, and 22 of the 219 embedding tests missed eager on CPU and CUDA: `embedding_int4wo_g128.decode_f32`, for one, is off by an RMSE of 1.8e-4 against a tolerance of 1e-5. The PTN carries the exact scales, so the engine was wrong, not the test, and a looser tolerance would hide that. Q4_0 can't take f32 scales without a ggml change, and dequantizing the table to f32 at load would match eager at about 7 times the memory.

A linear's matmul reads the scales from inside the blocks, but an embedding only gathers rows and can scale them afterwards. An int4 constant that only `aten.embedding` reads, with a scale that an fp16 round trip changes, now loads as Q4_0 blocks with unit scales, next to an f32 tensor of its group scales. `aten.embedding` gathers rows from both and multiplies them, which matches eager exactly. The f32 scales cost 1 bit per weight at group size 32 and a quarter bit at 128.

Every other int4 weight loads as before, with fp16 scales in its blocks:
- a weight shared with a linear, as in tied embeddings;
- a mutable weight, because an op can rewrite its blocks;
- a weight whose scales are all fp16 values, so most bf16 tables stay a single tensor.

`q4_affine` now also rejects zero-width int4 weights, which used to fail later as an allocation error. ggml itself is unchanged.

Both ggml-CPU CI targets now run the 22 int4 weight-only f32 embedding tests that the previous diff excluded.

Authored with Claude Code.

Differential Revision: [D130869557](https://our.internmc.facebook.com/intern/diff/D130869557/)

**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D130869557/)!
ghstack-source-id: 472753706