compile: output_shapes overrides for Slice/Scan/SDPA (unblocks shapeless whole-step compile)

[jonathan308]

mx.compile(f, shapeless=True) replaces shape-specialized retracing with shape-generic replay, but compile_replace needs primitive().output_shapes(real_inputs) per tape node and several primitives throw — the first being Slice ([Primitive::output_shapes] Slice cannot infer output shapes.). This blocks compiling an LLM decode step as one unit (the KV cache slice changes shape every token → per-token retrace today).

This PR adds the missing overrides needed by a decode-step graph:

• Slice::output_shapes — replicates the normalize_slice arithmetic from ops.cpp exactly (normalized start, unnormalized end, normalized strides); exact for cache-growth replays (shape only grows, positive starts). Documented limitation: slices created with negative start replay with the trace-time-baked value (pre-existing storage design; out of scope). Negative-strided slices with positive starts recompute exactly.
• Scan::output_shapes — shape-preserving (cumsum & co.); needed by the decode-mask pattern.
• ScaledDotProductAttention::output_shapes — was DEFINE_INPUT_OUTPUT_SHAPE() (= q shape; wrong when head_dim(v) != head_dim(qk), and missing the logsumexp second output). Now q.shape[:-1] + v.shape[-1:] plus the [..., 1] LSE output when enabled — matches the call-site construction in fast.cpp.

SliceUpdate, RMSNorm, RoPE already have correct overrides on main — verified by auditing all 128 primitive classes; no change.

Acceptance (Qwen3-0.6B-4bit decode step: offset-as-array, slice_update writes, full-buffer attention + cumsum mask, inputs=/outputs= state threading; 230-token prefill + 40 decode steps crossing the 256→512 cache-growth boundary)

• shapeless=True: 1 trace total, zero recompiles across growth, 40/40 tokens bit-identical to eager. (Pre-patch: throws on first compiled call.)
• shapeless=False: 2 traces (1 per growth), 40/40 bit-identical — unchanged behavior.

Two latent trace-time shape-baking traps were found and documented during acceptance (omitted-stop slices bake the trace-time axis size; broadcast_to of an under-sized attention mask bakes a Broadcast node that silently corrupts post-growth shapes in fused regions) — the acceptance recipe avoids both; fixing the defaults is out of scope (needs sentinel defaults in indexing.cpp).

Safety

The patch only adds shape-inference code paths used during compile(shapeless=True) replay; no eval kernel, op semantics, or eager path is touched. python/tests/test_compile.py + test_fast.py + test_ops.py: 254 passed, 327 subtests passed, 0 failed; 6 new tests added (slice arithmetic, negative strides, cumsum, decode-mask, sdpa incl. mismatched v head-dim).

Scope note

This is deliberately an enabler with a correctness contract, not a perf claim: on M3 Ultra, compiled decode is currently speed-neutral (GPU-bound), but shapeless compile removes per-256-token retraces and is a prerequisite for future graph-replay work. Follow-ups (not in this PR): Pad, Scatter, Split, AsStrided, View, Depends overrides; VJP primitives remain unsupported under shapeless.