Odds and ends: memory controls, contiguous, optional rms_norm weight

[dcvz]

Closes the pre-bump feature gaps flagged by community PRs, through the conformance pipeline:

• mlx_rs::memory — the complete 9-function allocator family (active/cache/peak observations, clear_cache, reset_peak, memory/cache/wired limits; setters return the previous value like the C API). Serial RAII-restoring integration tests; no oracle corpus since the figures are machine-dependent. Supersedes feat: add memory::set_memory_limit and memory::get_memory_limit #346, thanks @dschulmeist.
• Array::contiguous() + ContiguousOptions — row-major materialization, and AsSliceError::NotContiguous finally points at a concrete remedy. Exact-bit conformance cases over transposed/sliced/broadcast views incl. signed zero and fixed NaN payloads. Supersedes the op half of feat: add contiguity check, contiguous() op, and core fixes #326, thanks @ymote.
• fast::rms_norm(x, weight: Option<&Array>, eps) — genuine absence (C passes the empty handle), nn::RmsNorm unchanged as weighted. New fast conformance suite with None/weighted/wrong-shape cases + mutations. Supersedes fast::rms_norm: accept Option<&Array> for weight #347, thanks @Ogilthorp3.
• Shifts (feat: add element-wise left_shift and right_shift ops #345) deliberately NOT shipped: MLX 0.32.2 performs no shift-count validation and both backends use literal C++ shift operators — invalid counts are UB and the observed behavior is inconsistent across dtypes (int32 wraps the count mod width, uint8 truncates). Blessing that into a typed API would freeze a backend accident. Designed, recorded in ROADMAP.md, blocked on upstream defining the behavior (report drafted).

Evidence: full suite 1080/0, FFI leak gate pass (0 bytes, memory test binary included), verify-ledger + verify-oracle-boundary pass, API baseline blessed.

🤖 Generated with Claude Code