pulse: LCM-merge stream-axis dims + slope-based per-pulse for Range

[JulienBalianSonos]

When two parallel pulse paths converge at an elementwise op, the merged tensor's stream-axis dim is the Least Common Multiple (LCM) of the inputs' per-pulse stream dims, not the NumPy-style broadcast. Two semantics get conflated otherwise:

• Tensor shape compatibility (must match at runtime): non-stream axes use NumPy Broadcast. Equal or one is 1.
• Pulse-divisibility (a scalar constraint on per-pulse cycle): on the stream axis, two paths with steady-state size K_a and K_b are compatible at any pulse that is a multiple of LCM(K_a, K_b).

PulseWrappingOp::pulsed_output_facts was using the typed op's output_facts for shape merging, which falls through to multi_broadcast and produces Broadcast([K_a, K_b]) on the stream axis when paths produce different per-pulse sizes (e.g. ConvTranspose with kernel > stride at steady-state vs initial-state phases). Downstream pulse-divisibility checks then bail.

Changes

• New TDim::pulse_lcm helper (pure-integer LCM; returns None for symbolic, caller falls back).
• PulseWrappingOp::pulsed_output_facts post-processes typed output_facts: when the stream-axis dim is a Broadcast, replace with LCM of all stream-bearing inputs' stream dims. Non-stream axes keep Broadcast semantics (correct for runtime shape compatibility).
• Range::pulsify (pulse/src/ops/array/range.rs): replace stream_dim.substitute(symbol, pulse).to_usize() (full evaluated length, includes constant tail from upstream overlap-add) with slope-based guess_slope(symbol).0 × pulse_int (steady-state increment per pulse). The substitute form was forcing per-pulse to the kernel-state size when downstream paths were producing the steady-state size, surfacing exactly as the Broadcast(stride, kernel) we were merging.
• TDim::guess_slope made pub (was pub(super)) so pulse op-pulsifiers outside tract-data can use it.

Module-level doc

pulse/src/lib.rs now documents the stream-axis vs non-stream-axis merge contract so future maintainers don't reintroduce the conflation.

Surfaced by

Pocket-TTS Mimi audio decode (upsample[stride=16, kernel=32] → linear → view → unbind → arange-based mask → SDPA → linear → conv). Pre-fix: failed at Pulsification requires pulse Broadcast(16, 32) to be a stride (1) multiple. Post-fix: pulsifies cleanly through optimize at pulse=1, 2, 4, 8.

Depends on

#2201 (tdim: PartialEq second-chance). Needed for the LayerNorm-after-ConvTranspose case where the GCD-factoring rewrite produces algebraically-equal-but-structurally-different canonical forms.

Tests

• pulse_lcm_basic in tract-data.
• test_pulse_meet_with_arange_branch_types_through in tract-pulse: minimal repro of the upsample-then-arange-mask pattern; pulsifies cleanly post-fix.
• 132 + 34 lib tests green; 55 core-proptest-pulse tests green (no regressions in existing pad+conv, deconv, conv+conv, etc. proptests).

Known limitations / follow-ups

• pulse_lcm returns None for symbolic operands; callers fall back to existing Broadcast behavior (still safe). A symbolic LCM is worth following up.
• Op-specific pulsifiers (cnn::conv, deconv, fft, etc.) that compute stream-axis sizes in their own pulsify functions don't go through PulseWrappingOp; they should follow the same slope-based convention. This PR fixes Range; quick audit of others recommended.
• A pre-existing tract bug exists in Deconv pulse runtime semantics (output diverges at chunk boundaries for kernel > stride). Exposed by tract compare --stream once the type-level fix in this PR is in. Filed as #2203.