Fix Downsample and Log crashing on MLX arrays

[cboulay]

Two nodes that advertise Array API support could not run on MLX at all. Both turned up while benchmarking this package against the rules in Awni Hannun's Writing Fast MLX guide — the "before" case simply raised.

Downsample

Selected the kept samples with a NumPy integer index array:

pub_samples = np.where(samples == 0)[0]
data_slice = pub_samples          # -> ValueError: Cannot index mlx array using the given type

The kept samples are always an arithmetic sequence — first at (-s_idx) % q, then every q — so this is now a strided slice. A view rather than a gather on every backend, and the per-message np.arange goes away too.

Guide rule: prefer take/slicing over fancy indexing.

Log(clip_zero=True)

Reached xp.isdtype and finfo.smallest_normal, neither of which MLX exposes. It also guarded the clip on a host read:

has_non_positive = bool(xp.any(data <= 0))   # full device round-trip, every message

That stalls the pipeline exactly where a lazy backend would otherwise be running ahead. Clipping unconditionally is one fused elementwise pass. The only behavioral difference is that positive subnormals are now also raised to smallest_normal when nothing was <= 0 — which is what clip_zero is for.

Guide rule: avoid accidental frequent evaluation.

Measurements

M4 Pro, MLX 0.31.2, streaming shapes, async_eval per message with one synchronize at the end (i.e. how a streaming graph actually pays for it):

	30x256	128x512	512x1024
Log(clip_zero=True)	4.71x	4.62x	6.89x
Downsample selection, q=2	3.45x	2.25x	3.07x
Downsample selection, q=8	1.91x	2.61x	2.93x

mx.take was also measured as a candidate for the downsample gather and was not better than the gather (0.85–1.14x); the slice is the win.

Also here

• util.array.np_finfo — resolves dtype limits by name against NumPy. Backends disagree on what their own finfo carries (MLX's has only min/max/eps), and the float formats are IEEE-identical, so one lookup serves all of them. Returns None for dtypes NumPy cannot identify (e.g. bfloat16) rather than guessing.
• benchmarks/benchmark_mlx_gist_lessons.py — A/B micro-benchmarks walking each guide rule against what this package ships. Several candidates in it are measured and rejected (see below); it is kept so the constants can be re-derived on other hardware.
• benchmarks/benchmark_mlx_end_to_end.py — whole-node throughput under jittered chunk lengths, reporting min-of-repeats. Run-to-run spread here is ~6%, larger than several effects under test, so the isolated micro-result does not always survive.

Candidates the benchmarks rejected, recorded so nobody re-litigates them: abs(x)**2 → real²+imag² in spectrum (0.68–0.89x, i.e. slower), x @ W → x @ W_contig.T (no effect at our shapes), and caching the per-launch coef/valid_length Metal kernel constants (no effect).

Tests

Regression tests for both crashes, each asserting MLX agrees sample-for-sample with the NumPy path across chunked streaming. 4061 pass.