M8: native conv via mlx::core::conv_general - #17

Lift Backend.conv/4 from the BinaryBackend 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.

Layout translation: MLX conv_general expects NHWC input and OHWI
weight; Nx delivers tensors in the caller's layout plus
input/kernel/output_permutation opts pointing at canonical NCHW/OIHW.
Compose the caller's permutations with the NCHW↔NHWC and OIHW↔OHWI
transposes, applying 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). Two ordered transposes rather
than one composed — MLX lazy-fuses them and the step-wise form is
obviously correct across rank 3, 4, and 5.

Integer inputs get cast to out.type via Native.astype before the
transpose chain (Nx.conv returns float but does not cast operands;
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.

Whisper full-checkpoint conformance: 403.8s → 1.8s (~224× speedup).
ViT full-checkpoint conformance: similar scale (1.9s end-to-end).

New oracle suite in test/emily/backend_conv_test.exs covers 1-D/2-D/
3-D, stride/:same/:valid/asymmetric padding, kernel and input
dilation, grouped and depthwise conv, all three permutation options
(independently and combined for NHWC end-to-end), and integer-input
coercion.

test/emily/backend_fallbacks_test.exs: dropped the obsolete
"conv routes through BinaryBackend" test; added a batch_group_size
> 1 case asserting the rare fallback still matches BinaryBackend.