M8: native conv via mlx::core::conv_general

[ausimian]

Summary

• Lift Backend.conv/4 from the via_binary fallback onto Native.conv_general (already bound to mlx::core::conv_general since M1). The fallback was correct but CPU-bound — ≥90% of ViT and Whisper forward-pass cost.
• Handle the layout mismatch (Nx NCHW/OIHW ↔ MLX NHWC/OHWI) by composing the caller's input_permutation, kernel_permutation, and output_permutation opts with the canonical-layout transposes. Apply the inverse of output_permutation on the way out (Nx delivers it in user→canonical form; see deps/nx/lib/nx/shape.ex:729-735).
• Cast integer operands to out.type via Native.astype before dispatching — Nx.conv returns float but does not cast inputs, and MLX conv is float-only. Same-type astype is elided by MLX.
• batch_group_size > 1 and complex-typed conv still route through via_binary — no MLX primitive, and neither appears in the pinned Bumblebee ref.

Speedups

Suite	Before (fallback)	After (native)	Factor
mix test --only whisper_full	403.8s	1.8s	~224×
mix test --only vit_full	tens of seconds	1.9s	~20–50×

Test plan

• mix precommit — compile warnings-as-errors, format, credo --strict, 198 tests + 56 properties + 1 doctest green
• mix test --only conformance — 16/16 tiny-random (DistilBERT, Qwen3, ViT, Whisper)
• mix test --only vit_full — ViT full-checkpoint
• mix test --only whisper_full — Whisper full-checkpoint
• New test/emily/backend_conv_test.exs oracle suite vs Nx.BinaryBackend:
  • 1-D, 2-D, 3-D conv
  • stride, :same / :valid / explicit asymmetric padding
  • kernel_dilation and input_dilation > 1
  • grouped and depthwise conv
  • all three permutation options (independently and combined NHWC end-to-end)
  • integer-input coercion
• test/emily/backend_fallbacks_test.exs — removed obsolete "conv routes through BinaryBackend" test, added batch_group_size > 1 fallback coverage