Lower indexed_put / indexed_add (scatter) in the native Expr compiler

[ausimian]

Nx.indexed_put and Nx.indexed_add (scatter) now lower under
compiler: Emily.Compiler, native: true instead of forcing a graceful
fallback. Same shape as the FFT / take_along_axis additions — the MLX
ops and eager NIFs already existed (Native.{scatter,scatter_add} →
mx::{scatter,scatter_add}); only the compiler path was missing.
Implementing the one shared scatter path closes both ops.

What surfaced this

Continuing to run the Whisper speech_to_text serving with native: true
logged a fallback on :indexed_put. With this change that path compiles
native for the common (MLX-scatter-compatible) index layout.

Changes

• c_src/emily/opcodes.hpp — add Scatter/ScatterAdd (92–93), bump
  kOpcodeCount to 94, dispatch to mx::scatter (overwrite) /
  mx::scatter_add (accumulate). operands [target, updates, idx0, ...];
  iattrs [[axes...]].
• lib/emily/ir.ex — add the opcodes; one lower_op clause for
  :indexed_put/:indexed_add, mirroring Emily.Backend's apply_scatter:
  reuse the existing gather index-split + scatter_gather_compatible?
  helpers, port updates_shape_for_scatter, reshape updates into MLX's
  layout, emit the scatter opcode, coerce to out.type.
• compiler_equivalence_test.exs — native-vs-evaluator bit-identical
  cases: 2-D grid put/add, duplicate-index accumulation, partial-axis
  (axes: [0]) whole-row writes.

Behaviour note

Like the native gather, only MLX-scatter-compatible index layouts
(trailing dim == #axes, ≥1 batch dim) lower; other layouts raise during
lowering → the evaluator handles them under native_fallback: :eval
(unchanged from today, where any indexed_put falls back). Strict
improvement for the common layout.

Verification

• New scatter cases pass under native: true (suite default
  native_fallback: :raise, so a fallback would raise — passing proves
  full native lowering). Opcode-parity test confirms the IR↔C++ lockstep
  (92–93 / kOpcodeCount 94).
• mix precommit green: compile (warnings-as-errors), format, credo
  --strict, docs, 689 tests, 0 failures.