Let each event of the fused EPG kernel pay only for what it does

[mcencini]

The fused forward EPG kernel (_epg_kernel) does less work per event, with unchanged results. Stacked on #28: the base is cuda-off-axis-buffers, so the diff shows only the three kernel commits.

What changes

• Each event does only its own work. Every program reads the same event, so a branch on the event's kind and action is taken by all programs alike and costs no divergence. A pulse that turns rotates the states, a recorded readout reads them, and a shift or spoil moves or clears them. No event computes another kind's work only to discard it with tl.where.
• Phase trig is taken once, at launch. RF-spoiled phases grow without bound, and every program used to take the accurate cosine and sine at every pulse and readout. simulate_into now takes them once per event, in double precision. The kernel turns them by the transmit field's phase and gets the double angle from their products.
• The flip angle uses one reduction. The flip angle times B1 is per atom, so it can't be tabulated at launch. One Cody–Waite reduction by a quarter turn plus the single-precision Cephes polynomials gives its sine and cosine to about an ulp. Previously two library calls each made their own reduction.

Measured

Kernel time alone, best of repeats, on an RTX 4060 laptop GPU. The stream is a 7,132-event spoiled stream captured from a 3D GRE design:

trains × atoms	before	branches	+ phase tables	+ flip sincos
690	22.4 ms	—	15.4 ms	15.5 ms
6,210	193 ms	138 ms	116 ms	99 ms

Above about 200 trains × atoms the kernel is throughput-bound, at about 4.4 ns per event per problem before this change. Below that, the per-event latency of one problem's chain sets a floor of about 0.86 µs per event.

Two variants made no difference in this regime, so they are not in this PR:

• more problems per program (_TILE_ELEMENTS 128/256/512: 30, 39 and 203 ms at 690 problems, against 22 ms);
• a software-pipelined event loop (tl.range(..., num_stages=2..4)).

Checks

• Agreement with the old kernel on the captured stream:
  • the branches alone: 1.3e-8 relative;
  • with the phase tables: 4.8e-8;
  • with the flip-angle sincos: 3.7e-6, from the polynomial against libdevice's accurate routines.
• Test suite: 498 tests pass in tests/sequence, covering CUDA parity, pools, exchange, simulation, SSFP readouts, train batching, features, diffusion, flow, washout, dynamic transmit, exact profile, lineshape, transmit array, steady state, settling and fixed point.
• Not run: test_against_epgpy.py is skipped here because epgpy isn't installed.

🤖 Generated with Claude Code