[inureyes]

Closes #2465

Summary

Two numeric-parity fixes in one PR per the issue: PhiMoE reads the checkpoint's norm-epsilon key, and the antialiased bicubic kernel uses ATen's antialiased coefficient.

What changed

• src/models/phimoe.rs: layer_norm_eps now carries #[serde(alias = "rms_norm_eps")]. Phi-3.5-MoE checkpoints spell it rms_norm_eps: 1e-5 (the key mlx-lm's phimoe.py reads); previously the checkpoint value was ignored and the LayerNorms ran at the 1e-6 default. The layer_norm_eps spelling keeps parsing and 1e-6 stands when neither key exists.
• src/vision/encoders/resample_matrix.rs: interp_matrix picks a = -0.5 for a cubic kernel whenever antialias is set (ATen's HelperInterpCubic::aa_filter, what mlx-vlm uses) and keeps Keys' -0.75 without it. Non-antialiased output is bit-identical (pinned bytes in test).
• tests/fixtures/deepseekocr_sam/: torch-generated reference for the AA cubic matrix (16 to 5, 16 to 8, 16 to 24) and the SAM grid resample, from generate_reference.py with torch 2.14.1+cpu and mlx-vlm 0.7.8 (73d366ec), which agree within 9.8e-07.

Deviation from the proposed scale > 1 gate

Torch 2.14.1 uses a = -0.5 for the antialiased kernel in both directions, not only when downsampling: at the 16 to 24 upsample -0.5 matches torch within 6.1e-07 while -0.75 misses by 3.6e-02; at 16 to 5, 1.0e-07 vs 1.1e-02. Implemented as -0.5 whenever a cubic kernel is antialiased; support still stretches only when downsampling. No production caller upsamples with antialias (SAM downsamples 1024 to 64), so no existing behavior changes beyond the AA-downsample path.

Checkpoint comparison

On models/mlx/phi-3.5-moe-4bit with the prompt "The capital of France is" and 16 greedy tokens, mlx-lm (CPU-only mlx 0.32.4 from the oracle venv) produces Paris.\n\nIncorrect grammar: Me and my friend is going to the while mlxcel (CUDA MLX pinned 243bdf09, GPU) produces Paris. It is not only the country's largest; both start Paris. and diverge at token 3. A control run of mlxcel with the checkpoint's rms_norm_eps stripped (the pre-fix 1e-6 default) produces the identical stream, so the epsilon change does not alter generation on this prompt and the divergence is the cross-device numeric path, recorded rather than claimed as token-exact.

Test plan

• cargo test --profile test-fast phimoe -- --test-threads=1 (15 passed)
• cargo test --profile test-fast resample_matrix -- --test-threads=1 (9 passed)
• cargo test --profile test-fast deepseekocr -- --test-threads=1 (22 passed)
• cargo fmt --all -- --check and cargo clippy --profile test-fast -- -D warnings (clean)