fix(models): PhiMoE norm epsilon key and SAM antialiased bicubic a #2465

Problem / Background

Two small numeric-parity gaps left over from merged model PRs. Implement both in one PR, run the verification pass once at the end, and keep pushes to a minimum.

1. PhiMoE (found in PR fix(models): port upstream numeric fixes for four model families #2457): phimoe::ModelArgs reads layer_norm_eps (default 1e-6), but Phi-3.5-MoE checkpoints spell it rms_norm_eps: 1e-5, so the checkpoint value is ignored and the norms run at the wrong epsilon.
2. SAM vision encoder (left bit-identical in PR fix(vision): port LFM2-VL resampling, Qwen3-VL Conv3d guard and pos_embed #2459): the antialiased bicubic resize uses a = -0.75, while ATen and mlx-vlm use a = -0.5 for the antialiased kernel.

Current Behavior

• Item 1: src/models/phimoe.rs:67 declares layer_norm_eps with default_layer_norm_eps (line 137); it feeds LayerNorm::new at lines 510, 520 and 610.
• Item 2: src/vision/encoders/resample_matrix.rs:70 has const A: f32 = -0.75 for ResampleKernel::Cubic (line 42). interp_matrix(in, out, kernel, antialias) also serves non-antialiased callers, where -0.75 is correct. The SAM call site is src/vision/encoders/deepseekocr_sam.rs:81 and :82. Other antialiased callers: kimi_vl_pos_emb.rs:171 and :173 use ResampleKernel::Triangle (not affected); deepseekocr_clip.rs:190 is plain cubic without antialias; nemotron_h_nano_omni.rs:56 is antialias=False. Re-grep interp_matrix before editing to confirm no other cubic antialiased caller exists.

Proposed Solution

1. Read the key mlx-lm phimoe.py uses for its norm epsilon and match it. Accept both spellings in ModelArgs (serde alias or explicit fallback) with the checkpoint's rms_norm_eps winning when present, keeping 1e-6 only when neither key exists.
2. Select a = -0.5 only when kernel == Cubic && antialias && scale > 1 (the stretched-support path), keeping -0.75 on the non-antialiased path. Verify against mlx-vlm and torch interpolate(..., mode="bicubic", antialias=True).

Scope

In scope: src/models/phimoe.rs, src/vision/encoders/resample_matrix.rs, and tests.

Out of scope: other vision encoders' non-antialiased resizes.

Implementation Notes

• Reuse: existing fixtures style in resample_matrix tests and phimoe config tests.
• Constraints: non-antialiased cubic output must stay bit-identical.
• Edge cases: upsampling with antialias=True (kernel support is not stretched; confirm which coefficient torch uses and mirror it), scale exactly 1.
• Error handling: none new.

Acceptance Criteria

• Item 1: a config test parsing {"rms_norm_eps": 1e-5} expects 1e-5 and one without either key expects 1e-6. Fails on current main for the first case.
• Item 1: a short greedy or logit-trace check on a local Phi-3.5-MoE checkpoint, compared against mlx-lm, is recorded in the PR.
• Item 2: a resize parity test against a fixture generated by torch interpolate(..., mode="bicubic", antialias=True) (or mlx-vlm) on a downsampling case, passing within a stated tolerance. Fails on current main.
• Item 2: a test showing the non-antialiased cubic matrix is unchanged.
• SAM's real preprocessing path (deepseekocr_sam.rs) uses the corrected matrix.

Verification

cargo test --profile test-fast phimoe -- --test-threads=1
cargo test --profile test-fast resample_matrix -- --test-threads=1
cargo test --profile test-fast deepseekocr -- --test-threads=1
cargo fmt --all -- --check
cargo clippy --profile test-fast -- -D warnings

Generate the torch fixture offline and commit only the small reference values.